1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/MemoryLocation.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/IRBuilder.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/InstIterator.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/MDBuilder.h"
42 #include "llvm/IR/NoFolder.h"
43 #include "llvm/IR/PatternMatch.h"
44 #include "llvm/IR/Statepoint.h"
45 #include "llvm/IR/ValueHandle.h"
46 #include "llvm/IR/ValueMap.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Target/TargetLowering.h"
52 #include "llvm/Target/TargetSubtargetInfo.h"
53 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
54 #include "llvm/Transforms/Utils/BuildLibCalls.h"
55 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
59 using namespace llvm::PatternMatch;
61 #define DEBUG_TYPE "codegenprepare"
63 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
64 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
65 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
66 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
68 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
70 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
71 "computations were sunk");
72 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
73 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
74 STATISTIC(NumAndsAdded,
75 "Number of and mask instructions added to form ext loads");
76 STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized");
77 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
78 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
79 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
80 STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches");
81 STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed");
83 static cl::opt<bool> DisableBranchOpts(
84 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
85 cl::desc("Disable branch optimizations in CodeGenPrepare"));
88 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false),
89 cl::desc("Disable GC optimizations in CodeGenPrepare"));
91 static cl::opt<bool> DisableSelectToBranch(
92 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
93 cl::desc("Disable select to branch conversion."));
95 static cl::opt<bool> AddrSinkUsingGEPs(
96 "addr-sink-using-gep", cl::Hidden, cl::init(false),
97 cl::desc("Address sinking in CGP using GEPs."));
99 static cl::opt<bool> EnableAndCmpSinking(
100 "enable-andcmp-sinking", cl::Hidden, cl::init(true),
101 cl::desc("Enable sinkinig and/cmp into branches."));
103 static cl::opt<bool> DisableStoreExtract(
104 "disable-cgp-store-extract", cl::Hidden, cl::init(false),
105 cl::desc("Disable store(extract) optimizations in CodeGenPrepare"));
107 static cl::opt<bool> StressStoreExtract(
108 "stress-cgp-store-extract", cl::Hidden, cl::init(false),
109 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare"));
111 static cl::opt<bool> DisableExtLdPromotion(
112 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
113 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in "
116 static cl::opt<bool> StressExtLdPromotion(
117 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false),
118 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) "
119 "optimization in CodeGenPrepare"));
122 typedef SmallPtrSet<Instruction *, 16> SetOfInstrs;
123 typedef PointerIntPair<Type *, 1, bool> TypeIsSExt;
124 typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy;
125 class TypePromotionTransaction;
127 class CodeGenPrepare : public FunctionPass {
128 const TargetMachine *TM;
129 const TargetLowering *TLI;
130 const TargetTransformInfo *TTI;
131 const TargetLibraryInfo *TLInfo;
133 /// As we scan instructions optimizing them, this is the next instruction
134 /// to optimize. Transforms that can invalidate this should update it.
135 BasicBlock::iterator CurInstIterator;
137 /// Keeps track of non-local addresses that have been sunk into a block.
138 /// This allows us to avoid inserting duplicate code for blocks with
139 /// multiple load/stores of the same address.
140 ValueMap<Value*, Value*> SunkAddrs;
142 /// Keeps track of all instructions inserted for the current function.
143 SetOfInstrs InsertedInsts;
144 /// Keeps track of the type of the related instruction before their
145 /// promotion for the current function.
146 InstrToOrigTy PromotedInsts;
148 /// True if CFG is modified in any way.
151 /// True if optimizing for size.
154 /// DataLayout for the Function being processed.
155 const DataLayout *DL;
158 static char ID; // Pass identification, replacement for typeid
159 explicit CodeGenPrepare(const TargetMachine *TM = nullptr)
160 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr), DL(nullptr) {
161 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
163 bool runOnFunction(Function &F) override;
165 const char *getPassName() const override { return "CodeGen Prepare"; }
167 void getAnalysisUsage(AnalysisUsage &AU) const override {
168 AU.addPreserved<DominatorTreeWrapperPass>();
169 AU.addRequired<TargetLibraryInfoWrapperPass>();
170 AU.addRequired<TargetTransformInfoWrapperPass>();
174 bool eliminateFallThrough(Function &F);
175 bool eliminateMostlyEmptyBlocks(Function &F);
176 bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
177 void eliminateMostlyEmptyBlock(BasicBlock *BB);
178 bool optimizeBlock(BasicBlock &BB, bool& ModifiedDT);
179 bool optimizeInst(Instruction *I, bool& ModifiedDT);
180 bool optimizeMemoryInst(Instruction *I, Value *Addr,
181 Type *AccessTy, unsigned AS);
182 bool optimizeInlineAsmInst(CallInst *CS);
183 bool optimizeCallInst(CallInst *CI, bool& ModifiedDT);
184 bool moveExtToFormExtLoad(Instruction *&I);
185 bool optimizeExtUses(Instruction *I);
186 bool optimizeLoadExt(LoadInst *I);
187 bool optimizeSelectInst(SelectInst *SI);
188 bool optimizeShuffleVectorInst(ShuffleVectorInst *SI);
189 bool optimizeSwitchInst(SwitchInst *CI);
190 bool optimizeExtractElementInst(Instruction *Inst);
191 bool dupRetToEnableTailCallOpts(BasicBlock *BB);
192 bool placeDbgValues(Function &F);
193 bool sinkAndCmp(Function &F);
194 bool extLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI,
196 const SmallVectorImpl<Instruction *> &Exts,
197 unsigned CreatedInstCost);
198 bool splitBranchCondition(Function &F);
199 bool simplifyOffsetableRelocate(Instruction &I);
200 void stripInvariantGroupMetadata(Instruction &I);
204 char CodeGenPrepare::ID = 0;
205 INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare",
206 "Optimize for code generation", false, false)
208 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
209 return new CodeGenPrepare(TM);
214 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal);
215 Value* GetUntaintedAddress(Value* CurrentAddress);
217 // The depth we trace down a variable to look for its dependence set.
218 const unsigned kDependenceDepth = 4;
220 // Recursively looks for variables that 'Val' depends on at the given depth
221 // 'Depth', and adds them in 'DepSet'. If 'InsertOnlyLeafNodes' is true, only
222 // inserts the leaf node values; otherwise, all visited nodes are included in
223 // 'DepSet'. Note that constants will be ignored.
224 template <typename SetType>
225 void recursivelyFindDependence(SetType* DepSet, Value* Val,
226 bool InsertOnlyLeafNodes = false,
227 unsigned Depth = kDependenceDepth) {
228 if (Val == nullptr) {
231 if (!InsertOnlyLeafNodes && !isa<Constant>(Val)) {
235 // Cannot go deeper. Insert the leaf nodes.
236 if (InsertOnlyLeafNodes && !isa<Constant>(Val)) {
242 // Go one step further to explore the dependence of the operands.
243 Instruction* I = nullptr;
244 if ((I = dyn_cast<Instruction>(Val))) {
245 if (isa<LoadInst>(I)) {
246 // A load is considerd the leaf load of the dependence tree. Done.
249 } else if (I->isBinaryOp()) {
250 BinaryOperator* I = dyn_cast<BinaryOperator>(Val);
251 Value *Op0 = I->getOperand(0), *Op1 = I->getOperand(1);
252 recursivelyFindDependence(DepSet, Op0, Depth - 1);
253 recursivelyFindDependence(DepSet, Op1, Depth - 1);
254 } else if (I->isCast()) {
255 Value* Op0 = I->getOperand(0);
256 recursivelyFindDependence(DepSet, Op0, Depth - 1);
257 } else if (I->getOpcode() == Instruction::Select) {
258 Value* Op0 = I->getOperand(0);
259 Value* Op1 = I->getOperand(1);
260 Value* Op2 = I->getOperand(2);
261 recursivelyFindDependence(DepSet, Op0, Depth - 1);
262 recursivelyFindDependence(DepSet, Op1, Depth - 1);
263 recursivelyFindDependence(DepSet, Op2, Depth - 1);
264 } else if (I->getOpcode() == Instruction::GetElementPtr) {
265 for (unsigned i = 0; i < I->getNumOperands(); i++) {
266 recursivelyFindDependence(DepSet, I->getOperand(i), Depth - 1);
268 } else if (I->getOpcode() == Instruction::Store) {
269 auto* SI = dyn_cast<StoreInst>(Val);
270 recursivelyFindDependence(DepSet, SI->getPointerOperand(), Depth - 1);
271 recursivelyFindDependence(DepSet, SI->getValueOperand(), Depth - 1);
273 Value* Op0 = nullptr;
274 Value* Op1 = nullptr;
275 switch (I->getOpcode()) {
276 case Instruction::ICmp:
277 case Instruction::FCmp: {
278 Op0 = I->getOperand(0);
279 Op1 = I->getOperand(1);
280 recursivelyFindDependence(DepSet, Op0, Depth - 1);
281 recursivelyFindDependence(DepSet, Op1, Depth - 1);
285 // Be conservative. Add it and be done with it.
291 } else if (isa<Constant>(Val)) {
292 // Not interested in constant values. Done.
295 // Be conservative. Add it and be done with it.
301 // Helper function to create a Cast instruction.
302 Value* createCast(IRBuilder<true, NoFolder>& Builder, Value* DepVal,
303 Type* TargetIntegerType) {
304 Instruction::CastOps CastOp = Instruction::BitCast;
305 switch (DepVal->getType()->getTypeID()) {
306 case Type::IntegerTyID: {
307 CastOp = Instruction::SExt;
310 case Type::FloatTyID:
311 case Type::DoubleTyID: {
312 CastOp = Instruction::FPToSI;
315 case Type::PointerTyID: {
316 CastOp = Instruction::PtrToInt;
322 return Builder.CreateCast(CastOp, DepVal, TargetIntegerType);
325 // Given a value, if it's a tainted address, this function returns the
326 // instruction that ORs the "dependence value" with the "original address".
327 // Otherwise, returns nullptr. This instruction is the first OR instruction
328 // where one of its operand is an AND instruction with an operand being 0.
330 // E.g., it returns '%4 = or i32 %3, %2' given 'CurrentAddress' is '%5'.
331 // %0 = load i32, i32* @y, align 4, !tbaa !1
332 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
333 // %1 = sext i1 %cmp to i32
334 // %2 = ptrtoint i32* @x to i32
335 // %3 = and i32 %1, 0
336 // %4 = or i32 %3, %2
337 // %5 = inttoptr i32 %4 to i32*
338 // store i32 1, i32* %5, align 4
339 Instruction* getOrAddress(Value* CurrentAddress) {
340 // Is it a cast from integer to pointer type.
341 Instruction* OrAddress = nullptr;
342 Instruction* AndDep = nullptr;
343 Instruction* CastToInt = nullptr;
344 Value* ActualAddress = nullptr;
345 Constant* ZeroConst = nullptr;
347 const Instruction* CastToPtr = dyn_cast<Instruction>(CurrentAddress);
348 if (CastToPtr && CastToPtr->getOpcode() == Instruction::IntToPtr) {
349 // Is it an OR instruction: %1 = or %and, %actualAddress.
350 if ((OrAddress = dyn_cast<Instruction>(CastToPtr->getOperand(0))) &&
351 OrAddress->getOpcode() == Instruction::Or) {
352 // The first operand should be and AND instruction.
353 AndDep = dyn_cast<Instruction>(OrAddress->getOperand(0));
354 if (AndDep && AndDep->getOpcode() == Instruction::And) {
355 // Also make sure its first operand of the "AND" is 0, or the "AND" is
356 // marked explicitly by "NoInstCombine".
357 if ((ZeroConst = dyn_cast<Constant>(AndDep->getOperand(1))) &&
358 ZeroConst->isNullValue()) {
364 // Looks like it's not been tainted.
368 // Given a value, if it's a tainted address, this function returns the
369 // instruction that taints the "dependence value". Otherwise, returns nullptr.
370 // This instruction is the last AND instruction where one of its operand is 0.
371 // E.g., it returns '%3' given 'CurrentAddress' is '%5'.
372 // %0 = load i32, i32* @y, align 4, !tbaa !1
373 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
374 // %1 = sext i1 %cmp to i32
375 // %2 = ptrtoint i32* @x to i32
376 // %3 = and i32 %1, 0
377 // %4 = or i32 %3, %2
378 // %5 = inttoptr i32 %4 to i32*
379 // store i32 1, i32* %5, align 4
380 Instruction* getAndDependence(Value* CurrentAddress) {
381 // If 'CurrentAddress' is tainted, get the OR instruction.
382 auto* OrAddress = getOrAddress(CurrentAddress);
383 if (OrAddress == nullptr) {
387 // No need to check the operands.
388 auto* AndDepInst = dyn_cast<Instruction>(OrAddress->getOperand(0));
393 // Given a value, if it's a tainted address, this function returns
394 // the "dependence value", which is the first operand in the AND instruction.
395 // E.g., it returns '%1' given 'CurrentAddress' is '%5'.
396 // %0 = load i32, i32* @y, align 4, !tbaa !1
397 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
398 // %1 = sext i1 %cmp to i32
399 // %2 = ptrtoint i32* @x to i32
400 // %3 = and i32 %1, 0
401 // %4 = or i32 %3, %2
402 // %5 = inttoptr i32 %4 to i32*
403 // store i32 1, i32* %5, align 4
404 Value* getDependence(Value* CurrentAddress) {
405 auto* AndInst = getAndDependence(CurrentAddress);
406 if (AndInst == nullptr) {
409 return AndInst->getOperand(0);
412 // Given an address that has been tainted, returns the only condition it depends
413 // on, if any; otherwise, returns nullptr.
414 Value* getConditionDependence(Value* Address) {
415 auto* Dep = getDependence(Address);
416 if (Dep == nullptr) {
417 // 'Address' has not been dependence-tainted.
421 Value* Operand = Dep;
423 auto* Inst = dyn_cast<Instruction>(Operand);
424 if (Inst == nullptr) {
425 // Non-instruction type does not have condition dependence.
428 if (Inst->getOpcode() == Instruction::ICmp) {
431 if (Inst->getNumOperands() != 1) {
434 Operand = Inst->getOperand(0);
440 // Conservatively decides whether the dependence set of 'Val1' includes the
441 // dependence set of 'Val2'. If 'ExpandSecondValue' is false, we do not expand
442 // 'Val2' and use that single value as its dependence set.
443 // If it returns true, it means the dependence set of 'Val1' includes that of
444 // 'Val2'; otherwise, it only means we cannot conclusively decide it.
445 bool dependenceSetInclusion(Value* Val1, Value* Val2,
446 int Val1ExpandLevel = 2 * kDependenceDepth,
447 int Val2ExpandLevel = kDependenceDepth) {
448 typedef SmallSet<Value*, 8> IncludingSet;
449 typedef SmallSet<Value*, 4> IncludedSet;
451 IncludingSet DepSet1;
453 // Look for more depths for the including set.
454 recursivelyFindDependence(&DepSet1, Val1, false /*Insert all visited nodes*/,
456 recursivelyFindDependence(&DepSet2, Val2, true /*Only insert leaf nodes*/,
459 auto set_inclusion = [](IncludingSet FullSet, IncludedSet Subset) {
460 for (auto* Dep : Subset) {
461 if (0 == FullSet.count(Dep)) {
467 bool inclusion = set_inclusion(DepSet1, DepSet2);
468 DEBUG(dbgs() << "[dependenceSetInclusion]: " << inclusion << "\n");
469 DEBUG(dbgs() << "Including set for: " << *Val1 << "\n");
470 DEBUG(for (const auto* Dep : DepSet1) { dbgs() << "\t\t" << *Dep << "\n"; });
471 DEBUG(dbgs() << "Included set for: " << *Val2 << "\n");
472 DEBUG(for (const auto* Dep : DepSet2) { dbgs() << "\t\t" << *Dep << "\n"; });
477 // Recursively iterates through the operands spawned from 'DepVal'. If there
478 // exists a single value that 'DepVal' only depends on, we call that value the
479 // root dependence of 'DepVal' and return it. Otherwise, return 'DepVal'.
480 Value* getRootDependence(Value* DepVal) {
481 SmallSet<Value*, 8> DepSet;
482 for (unsigned depth = kDependenceDepth; depth > 0; --depth) {
483 recursivelyFindDependence(&DepSet, DepVal, true /*Only insert leaf nodes*/,
485 if (DepSet.size() == 1) {
486 return *DepSet.begin();
493 // This function actually taints 'DepVal' to the address to 'SI'. If the
495 // of 'SI' already depends on whatever 'DepVal' depends on, this function
496 // doesn't do anything and returns false. Otherwise, returns true.
498 // This effect forces the store and any stores that comes later to depend on
499 // 'DepVal'. For example, we have a condition "cond", and a store instruction
500 // "s: STORE addr, val". If we want "s" (and any later store) to depend on
501 // "cond", we do the following:
502 // %conv = sext i1 %cond to i32
503 // %addrVal = ptrtoint i32* %addr to i32
504 // %andCond = and i32 conv, 0;
505 // %orAddr = or i32 %andCond, %addrVal;
506 // %NewAddr = inttoptr i32 %orAddr to i32*;
508 // This is a more concrete example:
510 // %0 = load i32, i32* @y, align 4, !tbaa !1
511 // %cmp = icmp ne i32 %0, 42 // <== this is like the condition
512 // %1 = sext i1 %cmp to i32
513 // %2 = ptrtoint i32* @x to i32
514 // %3 = and i32 %1, 0
515 // %4 = or i32 %3, %2
516 // %5 = inttoptr i32 %4 to i32*
517 // store i32 1, i32* %5, align 4
518 bool taintStoreAddress(StoreInst* SI, Value* DepVal,
519 const char* calling_func = __builtin_FUNCTION()) {
520 DEBUG(dbgs() << "Called from " << calling_func << '\n');
521 // Set the insertion point right after the 'DepVal'.
522 Instruction* Inst = nullptr;
523 IRBuilder<true, NoFolder> Builder(SI);
524 BasicBlock* BB = SI->getParent();
525 Value* Address = SI->getPointerOperand();
526 Type* TargetIntegerType =
527 IntegerType::get(Address->getContext(),
528 BB->getModule()->getDataLayout().getPointerSizeInBits());
530 // Does SI's address already depends on whatever 'DepVal' depends on?
531 if (StoreAddressDependOnValue(SI, DepVal)) {
535 // Figure out if there's a root variable 'DepVal' depends on. For example, we
536 // can extract "getelementptr inbounds %struct, %struct* %0, i64 0, i32 123"
537 // to be "%struct* %0" since all other operands are constant.
538 DepVal = getRootDependence(DepVal);
540 // Is this already a dependence-tainted store?
541 Value* OldDep = getDependence(Address);
543 // The address of 'SI' has already been tainted. Just need to absorb the
544 // DepVal to the existing dependence in the address of SI.
545 Instruction* AndDep = getAndDependence(Address);
546 IRBuilder<true, NoFolder> Builder(AndDep);
547 Value* NewDep = nullptr;
548 if (DepVal->getType() == AndDep->getType()) {
549 NewDep = Builder.CreateAnd(OldDep, DepVal);
551 NewDep = Builder.CreateAnd(
552 OldDep, createCast(Builder, DepVal, TargetIntegerType));
555 auto* NewDepInst = dyn_cast<Instruction>(NewDep);
557 // Use the new AND instruction as the dependence
558 AndDep->setOperand(0, NewDep);
562 // SI's address has not been tainted. Now taint it with 'DepVal'.
563 Value* CastDepToInt = createCast(Builder, DepVal, TargetIntegerType);
564 Value* PtrToIntCast = Builder.CreatePtrToInt(Address, TargetIntegerType);
566 Builder.CreateAnd(CastDepToInt, ConstantInt::get(TargetIntegerType, 0));
567 auto AndInst = dyn_cast<Instruction>(AndDepVal);
568 // XXX-comment: The original IR InstCombiner would change our and instruction
569 // to a select and then the back end optimize the condition out. We attach a
570 // flag to instructions and set it here to inform the InstCombiner to not to
571 // touch this and instruction at all.
572 Value* OrAddr = Builder.CreateOr(AndDepVal, PtrToIntCast);
573 Value* NewAddr = Builder.CreateIntToPtr(OrAddr, Address->getType());
575 DEBUG(dbgs() << "[taintStoreAddress]\n"
576 << "Original store: " << *SI << '\n');
577 SI->setOperand(1, NewAddr);
580 DEBUG(dbgs() << "\tTargetIntegerType: " << *TargetIntegerType << '\n'
581 << "\tCast dependence value to integer: " << *CastDepToInt
583 << "\tCast address to integer: " << *PtrToIntCast << '\n'
584 << "\tAnd dependence value: " << *AndDepVal << '\n'
585 << "\tOr address: " << *OrAddr << '\n'
586 << "\tCast or instruction to address: " << *NewAddr << "\n\n");
591 // Looks for the previous store in the if block --- 'BrBB', which makes the
592 // speculative store 'StoreToHoist' safe.
593 Value* getSpeculativeStoreInPrevBB(StoreInst* StoreToHoist, BasicBlock* BrBB) {
594 assert(StoreToHoist && "StoreToHoist must be a real store");
596 Value* StorePtr = StoreToHoist->getPointerOperand();
598 // Look for a store to the same pointer in BrBB.
599 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(), RE = BrBB->rend();
601 Instruction* CurI = &*RI;
603 StoreInst* SI = dyn_cast<StoreInst>(CurI);
604 // Found the previous store make sure it stores to the same location.
605 // XXX-update: If the previous store's original untainted address are the
606 // same as 'StorePtr', we are also good to hoist the store.
607 if (SI && (SI->getPointerOperand() == StorePtr ||
608 GetUntaintedAddress(SI->getPointerOperand()) == StorePtr)) {
609 // Found the previous store, return its value operand.
615 "We should not reach here since this store is safe to speculate");
618 // XXX-comment: Returns true if it changes the code, false otherwise (the branch
619 // condition already depends on 'DepVal'.
620 bool taintConditionalBranch(BranchInst* BI, Value* DepVal) {
621 assert(BI->isConditional());
622 auto* Cond = BI->getOperand(0);
623 if (dependenceSetInclusion(Cond, DepVal)) {
624 // The dependence/ordering is self-evident.
628 IRBuilder<true, NoFolder> Builder(BI);
630 Builder.CreateAnd(DepVal, ConstantInt::get(DepVal->getType(), 0));
632 Builder.CreateTrunc(AndDep, IntegerType::get(DepVal->getContext(), 1));
633 auto* OrCond = Builder.CreateOr(TruncAndDep, Cond);
634 BI->setOperand(0, OrCond);
637 DEBUG(dbgs() << "\tTainted branch condition:\n" << *BI->getParent());
642 bool ConditionalBranchDependsOnValue(BranchInst* BI, Value* DepVal) {
643 assert(BI->isConditional());
644 auto* Cond = BI->getOperand(0);
645 return dependenceSetInclusion(Cond, DepVal);
648 // XXX-update: For a relaxed load 'LI', find the first immediate atomic store or
649 // the first conditional branch. Returns nullptr if there's no such immediately
650 // following store/branch instructions, which we can only enforce the load with
652 Instruction* findFirstStoreCondBranchInst(LoadInst* LI) {
653 // In some situations, relaxed loads can be left as is:
654 // 1. The relaxed load is used to calculate the address of the immediate
656 // 2. The relaxed load is used as a condition in the immediate following
657 // condition, and there are no stores in between. This is actually quite
659 // int r1 = x.load(relaxed);
661 // y.store(1, relaxed);
663 // However, in this function, we don't deal with them directly. Instead, we
664 // just find the immediate following store/condition branch and return it.
666 auto* BB = LI->getParent();
668 auto BBI = BasicBlock::iterator(LI);
671 for (; BBI != BE; BBI++) {
672 auto* Inst = dyn_cast<Instruction>(&*BBI);
673 if (Inst == nullptr) {
676 if (Inst->getOpcode() == Instruction::Store) {
678 } else if (Inst->getOpcode() == Instruction::Br) {
679 auto* BrInst = dyn_cast<BranchInst>(Inst);
680 if (BrInst->isConditional()) {
683 // Reinitialize iterators with the destination of the unconditional
685 BB = BrInst->getSuccessor(0);
698 // XXX-comment: Returns whether the code has been changed.
699 bool taintMonotonicLoads(const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
700 bool Changed = false;
701 for (auto* LI : MonotonicLoadInsts) {
702 auto* FirstInst = findFirstStoreCondBranchInst(LI);
703 if (FirstInst == nullptr) {
704 // We don't seem to be able to taint a following store/conditional branch
705 // instruction. Simply make it acquire.
706 DEBUG(dbgs() << "[RelaxedLoad]: Transformed to acquire load\n"
708 LI->setOrdering(Acquire);
712 // Taint 'FirstInst', which could be a store or a condition branch
714 if (FirstInst->getOpcode() == Instruction::Store) {
715 Changed |= taintStoreAddress(dyn_cast<StoreInst>(FirstInst), LI);
716 } else if (FirstInst->getOpcode() == Instruction::Br) {
717 Changed |= taintConditionalBranch(dyn_cast<BranchInst>(FirstInst), LI);
719 assert(false && "findFirstStoreCondBranchInst() should return a "
720 "store/condition branch instruction");
726 // Inserts a fake conditional branch right after the instruction 'SplitInst',
727 // and the branch condition is 'Condition'. 'SplitInst' will be placed in the
728 // newly created block.
729 void AddFakeConditionalBranch(Instruction* SplitInst, Value* Condition) {
730 auto* BB = SplitInst->getParent();
731 TerminatorInst* ThenTerm = nullptr;
732 TerminatorInst* ElseTerm = nullptr;
733 SplitBlockAndInsertIfThenElse(Condition, SplitInst, &ThenTerm, &ElseTerm);
734 assert(ThenTerm && ElseTerm &&
735 "Then/Else terminators cannot be empty after basic block spliting");
736 auto* ThenBB = ThenTerm->getParent();
737 auto* ElseBB = ElseTerm->getParent();
738 auto* TailBB = ThenBB->getSingleSuccessor();
739 assert(TailBB && "Tail block cannot be empty after basic block spliting");
741 ThenBB->disableCanEliminateBlock();
742 ThenBB->disableCanEliminateBlock();
743 TailBB->disableCanEliminateBlock();
744 ThenBB->setName(BB->getName() + "Then.Fake");
745 ElseBB->setName(BB->getName() + "Else.Fake");
746 DEBUG(dbgs() << "Add fake conditional branch:\n"
748 << *ThenBB << "Else Block:\n"
752 // Returns true if the code is changed, and false otherwise.
753 void TaintRelaxedLoads(LoadInst* LI) {
754 // For better performance, we can add a "AND X 0" instruction before the
756 auto* FirstInst = findFirstStoreCondBranchInst(LI);
757 Instruction* InsertPoint = nullptr;
758 if (FirstInst == nullptr) {
759 InsertPoint = LI->getParent()->getTerminator();
760 InsertPoint = LI->getNextNode();
762 InsertPoint = LI->getNextNode();
764 IRBuilder<true, NoFolder> Builder(InsertPoint);
765 auto* AndZero = dyn_cast<Instruction>(
766 Builder.CreateAnd(LI, Constant::getNullValue(LI->getType())));
767 auto* FakeCondition = dyn_cast<Instruction>(Builder.CreateICmp(
768 CmpInst::ICMP_NE, AndZero, Constant::getNullValue(LI->getType())));
769 AddFakeConditionalBranch(FakeCondition->getNextNode(), FakeCondition);
772 // XXX-comment: Returns whether the code has been changed.
773 bool AddFakeConditionalBranchAfterMonotonicLoads(
774 const SmallVector<LoadInst*, 1>& MonotonicLoadInsts) {
775 bool Changed = false;
776 for (auto* LI : MonotonicLoadInsts) {
777 auto* FirstInst = findFirstStoreCondBranchInst(LI);
778 if (FirstInst != nullptr) {
779 if (FirstInst->getOpcode() == Instruction::Store) {
780 if (StoreAddressDependOnValue(dyn_cast<StoreInst>(FirstInst), LI)) {
783 } else if (FirstInst->getOpcode() == Instruction::Br) {
784 if (ConditionalBranchDependsOnValue(dyn_cast<BranchInst>(FirstInst),
789 dbgs() << "FirstInst=" << *FirstInst << "\n";
790 assert(false && "findFirstStoreCondBranchInst() should return a "
791 "store/condition branch instruction");
795 // We really need to process the relaxed load now.
796 StoreInst* SI = nullptr;;
797 if (FirstInst && (SI = dyn_cast<StoreInst>(FirstInst))) {
798 // For immediately coming stores, taint the address of the store.
799 taintStoreAddress(SI, LI);
801 // For immediately coming branch, directly add a fake branch.
802 TaintRelaxedLoads(LI);
809 /**** Implementations of public methods for dependence tainting ****/
810 Value* GetUntaintedAddress(Value* CurrentAddress) {
811 auto* OrAddress = getOrAddress(CurrentAddress);
812 if (OrAddress == nullptr) {
813 // Is it tainted by a select instruction?
814 auto* Inst = dyn_cast<Instruction>(CurrentAddress);
815 if (nullptr != Inst && Inst->getOpcode() == Instruction::Select) {
816 // A selection instruction.
817 if (Inst->getOperand(1) == Inst->getOperand(2)) {
818 return Inst->getOperand(1);
822 return CurrentAddress;
824 Value* ActualAddress = nullptr;
826 auto* CastToInt = dyn_cast<Instruction>(OrAddress->getOperand(1));
827 if (CastToInt && CastToInt->getOpcode() == Instruction::PtrToInt) {
828 return CastToInt->getOperand(0);
830 // This should be a IntToPtr constant expression.
831 ConstantExpr* PtrToIntExpr =
832 dyn_cast<ConstantExpr>(OrAddress->getOperand(1));
833 if (PtrToIntExpr && PtrToIntExpr->getOpcode() == Instruction::PtrToInt) {
834 return PtrToIntExpr->getOperand(0);
838 // Looks like it's not been dependence-tainted. Returns itself.
839 return CurrentAddress;
842 MemoryLocation GetUntaintedMemoryLocation(StoreInst* SI) {
844 SI->getAAMetadata(AATags);
845 const auto& DL = SI->getModule()->getDataLayout();
846 const auto* OriginalAddr = GetUntaintedAddress(SI->getPointerOperand());
847 DEBUG(if (OriginalAddr != SI->getPointerOperand()) {
848 dbgs() << "[GetUntaintedMemoryLocation]\n"
849 << "Storing address: " << *SI->getPointerOperand()
850 << "\nUntainted address: " << *OriginalAddr << "\n";
852 return MemoryLocation(OriginalAddr,
853 DL.getTypeStoreSize(SI->getValueOperand()->getType()),
857 bool TaintDependenceToStore(StoreInst* SI, Value* DepVal) {
858 if (dependenceSetInclusion(SI, DepVal)) {
862 bool tainted = taintStoreAddress(SI, DepVal);
867 bool TaintDependenceToStoreAddress(StoreInst* SI, Value* DepVal) {
868 if (dependenceSetInclusion(SI->getPointerOperand(), DepVal)) {
872 bool tainted = taintStoreAddress(SI, DepVal);
877 bool CompressTaintedStore(BasicBlock* BB) {
878 // This function looks for windows of adajcent stores in 'BB' that satisfy the
879 // following condition (and then do optimization):
880 // *Addr(d1) = v1, d1 is a condition and is the only dependence the store's
881 // address depends on && Dep(v1) includes Dep(d1);
882 // *Addr(d2) = v2, d2 is a condition and is the only dependnece the store's
883 // address depends on && Dep(v2) includes Dep(d2) &&
884 // Dep(d2) includes Dep(d1);
886 // *Addr(dN) = vN, dN is a condition and is the only dependence the store's
887 // address depends on && Dep(dN) includes Dep(d"N-1").
889 // As a result, Dep(dN) includes [Dep(d1) V ... V Dep(d"N-1")], so we can
890 // safely transform the above to the following. In between these stores, we
891 // can omit untainted stores to the same address 'Addr' since they internally
892 // have dependence on the previous stores on the same address.
897 for (auto BI = BB->begin(), BE = BB->end(); BI != BE; BI++) {
898 // Look for the first store in such a window of adajacent stores.
899 auto* FirstSI = dyn_cast<StoreInst>(&*BI);
904 // The first store in the window must be tainted.
905 auto* UntaintedAddress = GetUntaintedAddress(FirstSI->getPointerOperand());
906 if (UntaintedAddress == FirstSI->getPointerOperand()) {
910 // The first store's address must directly depend on and only depend on a
912 auto* FirstSIDepCond = getConditionDependence(FirstSI->getPointerOperand());
913 if (nullptr == FirstSIDepCond) {
917 // Dep(first store's storing value) includes Dep(tainted dependence).
918 if (!dependenceSetInclusion(FirstSI->getValueOperand(), FirstSIDepCond)) {
922 // Look for subsequent stores to the same address that satisfy the condition
923 // of "compressing the dependence".
924 SmallVector<StoreInst*, 8> AdajacentStores;
925 AdajacentStores.push_back(FirstSI);
926 auto BII = BasicBlock::iterator(FirstSI);
927 for (BII++; BII != BE; BII++) {
928 auto* CurrSI = dyn_cast<StoreInst>(&*BII);
930 if (BII->mayHaveSideEffects()) {
931 // Be conservative. Instructions with side effects are similar to
938 auto* OrigAddress = GetUntaintedAddress(CurrSI->getPointerOperand());
939 auto* CurrSIDepCond = getConditionDependence(CurrSI->getPointerOperand());
940 // All other stores must satisfy either:
941 // A. 'CurrSI' is an untainted store to the same address, or
942 // B. the combination of the following 5 subconditions:
944 // 2. Untainted address is the same as the group's address;
945 // 3. The address is tainted with a sole value which is a condition;
946 // 4. The storing value depends on the condition in 3.
947 // 5. The condition in 3 depends on the previous stores dependence
950 // Condition A. Should ignore this store directly.
951 if (OrigAddress == CurrSI->getPointerOperand() &&
952 OrigAddress == UntaintedAddress) {
955 // Check condition B.
956 Value* Cond = nullptr;
957 if (OrigAddress == CurrSI->getPointerOperand() ||
958 OrigAddress != UntaintedAddress || CurrSIDepCond == nullptr ||
959 !dependenceSetInclusion(CurrSI->getValueOperand(), CurrSIDepCond)) {
960 // Check condition 1, 2, 3 & 4.
964 // Check condition 5.
965 StoreInst* PrevSI = AdajacentStores[AdajacentStores.size() - 1];
966 auto* PrevSIDepCond = getConditionDependence(PrevSI->getPointerOperand());
967 assert(PrevSIDepCond &&
968 "Store in the group must already depend on a condtion");
969 if (!dependenceSetInclusion(CurrSIDepCond, PrevSIDepCond)) {
973 AdajacentStores.push_back(CurrSI);
976 if (AdajacentStores.size() == 1) {
977 // The outer loop should keep looking from the next store.
981 // Now we have such a group of tainted stores to the same address.
982 DEBUG(dbgs() << "[CompressTaintedStore]\n");
983 DEBUG(dbgs() << "Original BB\n");
984 DEBUG(dbgs() << *BB << '\n');
985 auto* LastSI = AdajacentStores[AdajacentStores.size() - 1];
986 for (unsigned i = 0; i < AdajacentStores.size() - 1; ++i) {
987 auto* SI = AdajacentStores[i];
989 // Use the original address for stores before the last one.
990 SI->setOperand(1, UntaintedAddress);
992 DEBUG(dbgs() << "Store address has been reversed: " << *SI << '\n';);
994 // XXX-comment: Try to make the last store use fewer registers.
995 // If LastSI's storing value is a select based on the condition with which
996 // its address is tainted, transform the tainted address to a select
997 // instruction, as follows:
998 // r1 = Select Cond ? A : B
1003 // r1 = Select Cond ? A : B
1004 // r2 = Select Cond ? Addr : Addr
1006 // The idea is that both Select instructions depend on the same condition,
1007 // so hopefully the backend can generate two cmov instructions for them (and
1008 // this saves the number of registers needed).
1009 auto* LastSIDep = getConditionDependence(LastSI->getPointerOperand());
1010 auto* LastSIValue = dyn_cast<Instruction>(LastSI->getValueOperand());
1011 if (LastSIValue && LastSIValue->getOpcode() == Instruction::Select &&
1012 LastSIValue->getOperand(0) == LastSIDep) {
1013 // XXX-comment: Maybe it's better for us to just leave it as an and/or
1014 // dependence pattern.
1016 IRBuilder<true, NoFolder> Builder(LastSI);
1018 Builder.CreateSelect(LastSIDep, UntaintedAddress, UntaintedAddress);
1019 LastSI->setOperand(1, Address);
1020 DEBUG(dbgs() << "The last store becomes :" << *LastSI << "\n\n";);
1028 bool PassDependenceToStore(Value* OldAddress, StoreInst* NewStore) {
1029 Value* OldDep = getDependence(OldAddress);
1030 // Return false when there's no dependence to pass from the OldAddress.
1035 // No need to pass the dependence to NewStore's address if it already depends
1036 // on whatever 'OldAddress' depends on.
1037 if (StoreAddressDependOnValue(NewStore, OldDep)) {
1040 return taintStoreAddress(NewStore, OldAddress);
1043 SmallSet<Value*, 8> FindDependence(Value* Val) {
1044 SmallSet<Value*, 8> DepSet;
1045 recursivelyFindDependence(&DepSet, Val, true /*Only insert leaf nodes*/);
1049 bool StoreAddressDependOnValue(StoreInst* SI, Value* DepVal) {
1050 return dependenceSetInclusion(SI->getPointerOperand(), DepVal);
1053 bool StoreDependOnValue(StoreInst* SI, Value* Dep) {
1054 return dependenceSetInclusion(SI, Dep);
1061 bool CodeGenPrepare::runOnFunction(Function &F) {
1062 bool EverMadeChange = false;
1064 if (skipOptnoneFunction(F))
1067 DL = &F.getParent()->getDataLayout();
1069 // Clear per function information.
1070 InsertedInsts.clear();
1071 PromotedInsts.clear();
1075 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
1076 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1077 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1078 OptSize = F.optForSize();
1080 /// This optimization identifies DIV instructions that can be
1081 /// profitably bypassed and carried out with a shorter, faster divide.
1082 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
1083 const DenseMap<unsigned int, unsigned int> &BypassWidths =
1084 TLI->getBypassSlowDivWidths();
1085 BasicBlock* BB = &*F.begin();
1086 while (BB != nullptr) {
1087 // bypassSlowDivision may create new BBs, but we don't want to reapply the
1088 // optimization to those blocks.
1089 BasicBlock* Next = BB->getNextNode();
1090 EverMadeChange |= bypassSlowDivision(BB, BypassWidths);
1095 // Eliminate blocks that contain only PHI nodes and an
1096 // unconditional branch.
1097 EverMadeChange |= eliminateMostlyEmptyBlocks(F);
1099 // llvm.dbg.value is far away from the value then iSel may not be able
1100 // handle it properly. iSel will drop llvm.dbg.value if it can not
1101 // find a node corresponding to the value.
1102 EverMadeChange |= placeDbgValues(F);
1104 // If there is a mask, compare against zero, and branch that can be combined
1105 // into a single target instruction, push the mask and compare into branch
1106 // users. Do this before OptimizeBlock -> OptimizeInst ->
1107 // OptimizeCmpExpression, which perturbs the pattern being searched for.
1108 if (!DisableBranchOpts) {
1109 EverMadeChange |= sinkAndCmp(F);
1110 EverMadeChange |= splitBranchCondition(F);
1113 bool MadeChange = true;
1114 while (MadeChange) {
1116 for (Function::iterator I = F.begin(); I != F.end(); ) {
1117 BasicBlock *BB = &*I++;
1118 bool ModifiedDTOnIteration = false;
1119 MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration);
1121 // Restart BB iteration if the dominator tree of the Function was changed
1122 if (ModifiedDTOnIteration)
1125 EverMadeChange |= MadeChange;
1130 if (!DisableBranchOpts) {
1132 SmallPtrSet<BasicBlock*, 8> WorkList;
1133 for (BasicBlock &BB : F) {
1134 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB));
1135 MadeChange |= ConstantFoldTerminator(&BB, true);
1136 if (!MadeChange) continue;
1138 for (SmallVectorImpl<BasicBlock*>::iterator
1139 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1140 if (pred_begin(*II) == pred_end(*II))
1141 WorkList.insert(*II);
1144 // Delete the dead blocks and any of their dead successors.
1145 MadeChange |= !WorkList.empty();
1146 while (!WorkList.empty()) {
1147 BasicBlock *BB = *WorkList.begin();
1149 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
1151 DeleteDeadBlock(BB);
1153 for (SmallVectorImpl<BasicBlock*>::iterator
1154 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
1155 if (pred_begin(*II) == pred_end(*II))
1156 WorkList.insert(*II);
1159 // Merge pairs of basic blocks with unconditional branches, connected by
1161 if (EverMadeChange || MadeChange)
1162 MadeChange |= eliminateFallThrough(F);
1164 EverMadeChange |= MadeChange;
1167 if (!DisableGCOpts) {
1168 SmallVector<Instruction *, 2> Statepoints;
1169 for (BasicBlock &BB : F)
1170 for (Instruction &I : BB)
1171 if (isStatepoint(I))
1172 Statepoints.push_back(&I);
1173 for (auto &I : Statepoints)
1174 EverMadeChange |= simplifyOffsetableRelocate(*I);
1177 // XXX-comment: Delay dealing with relaxed loads in this function to avoid
1178 // further changes done by other passes (e.g., SimplifyCFG).
1179 // Collect all the relaxed loads.
1180 SmallVector<LoadInst*, 1> MonotonicLoadInsts;
1181 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
1182 if (I->isAtomic()) {
1183 switch (I->getOpcode()) {
1184 case Instruction::Load: {
1185 auto* LI = dyn_cast<LoadInst>(&*I);
1186 if (LI->getOrdering() == Monotonic) {
1187 MonotonicLoadInsts.push_back(LI);
1198 AddFakeConditionalBranchAfterMonotonicLoads(MonotonicLoadInsts);
1200 return EverMadeChange;
1203 /// Merge basic blocks which are connected by a single edge, where one of the
1204 /// basic blocks has a single successor pointing to the other basic block,
1205 /// which has a single predecessor.
1206 bool CodeGenPrepare::eliminateFallThrough(Function &F) {
1207 bool Changed = false;
1208 // Scan all of the blocks in the function, except for the entry block.
1209 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1210 BasicBlock *BB = &*I++;
1211 // If the destination block has a single pred, then this is a trivial
1212 // edge, just collapse it.
1213 BasicBlock *SinglePred = BB->getSinglePredecessor();
1215 // Don't merge if BB's address is taken.
1216 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
1218 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
1219 if (Term && !Term->isConditional()) {
1221 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
1222 // Remember if SinglePred was the entry block of the function.
1223 // If so, we will need to move BB back to the entry position.
1224 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1225 MergeBasicBlockIntoOnlyPred(BB, nullptr);
1227 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1228 BB->moveBefore(&BB->getParent()->getEntryBlock());
1230 // We have erased a block. Update the iterator.
1231 I = BB->getIterator();
1237 /// Eliminate blocks that contain only PHI nodes, debug info directives, and an
1238 /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split
1239 /// edges in ways that are non-optimal for isel. Start by eliminating these
1240 /// blocks so we can split them the way we want them.
1241 bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) {
1242 bool MadeChange = false;
1243 // Note that this intentionally skips the entry block.
1244 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) {
1245 BasicBlock *BB = &*I++;
1246 // If this block doesn't end with an uncond branch, ignore it.
1247 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
1248 if (!BI || !BI->isUnconditional())
1251 // If the instruction before the branch (skipping debug info) isn't a phi
1252 // node, then other stuff is happening here.
1253 BasicBlock::iterator BBI = BI->getIterator();
1254 if (BBI != BB->begin()) {
1256 while (isa<DbgInfoIntrinsic>(BBI)) {
1257 if (BBI == BB->begin())
1261 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
1265 // Do not break infinite loops.
1266 BasicBlock *DestBB = BI->getSuccessor(0);
1270 if (!canMergeBlocks(BB, DestBB))
1273 eliminateMostlyEmptyBlock(BB);
1279 /// Return true if we can merge BB into DestBB if there is a single
1280 /// unconditional branch between them, and BB contains no other non-phi
1282 bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB,
1283 const BasicBlock *DestBB) const {
1284 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
1285 // the successor. If there are more complex condition (e.g. preheaders),
1286 // don't mess around with them.
1287 BasicBlock::const_iterator BBI = BB->begin();
1288 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1289 for (const User *U : PN->users()) {
1290 const Instruction *UI = cast<Instruction>(U);
1291 if (UI->getParent() != DestBB || !isa<PHINode>(UI))
1293 // If User is inside DestBB block and it is a PHINode then check
1294 // incoming value. If incoming value is not from BB then this is
1295 // a complex condition (e.g. preheaders) we want to avoid here.
1296 if (UI->getParent() == DestBB) {
1297 if (const PHINode *UPN = dyn_cast<PHINode>(UI))
1298 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
1299 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
1300 if (Insn && Insn->getParent() == BB &&
1301 Insn->getParent() != UPN->getIncomingBlock(I))
1308 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
1309 // and DestBB may have conflicting incoming values for the block. If so, we
1310 // can't merge the block.
1311 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
1312 if (!DestBBPN) return true; // no conflict.
1314 // Collect the preds of BB.
1315 SmallPtrSet<const BasicBlock*, 16> BBPreds;
1316 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1317 // It is faster to get preds from a PHI than with pred_iterator.
1318 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1319 BBPreds.insert(BBPN->getIncomingBlock(i));
1321 BBPreds.insert(pred_begin(BB), pred_end(BB));
1324 // Walk the preds of DestBB.
1325 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
1326 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
1327 if (BBPreds.count(Pred)) { // Common predecessor?
1328 BBI = DestBB->begin();
1329 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
1330 const Value *V1 = PN->getIncomingValueForBlock(Pred);
1331 const Value *V2 = PN->getIncomingValueForBlock(BB);
1333 // If V2 is a phi node in BB, look up what the mapped value will be.
1334 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
1335 if (V2PN->getParent() == BB)
1336 V2 = V2PN->getIncomingValueForBlock(Pred);
1338 // If there is a conflict, bail out.
1339 if (V1 != V2) return false;
1348 /// Eliminate a basic block that has only phi's and an unconditional branch in
1350 void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) {
1351 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1352 BasicBlock *DestBB = BI->getSuccessor(0);
1354 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
1356 // If the destination block has a single pred, then this is a trivial edge,
1357 // just collapse it.
1358 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
1359 if (SinglePred != DestBB) {
1360 // Remember if SinglePred was the entry block of the function. If so, we
1361 // will need to move BB back to the entry position.
1362 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
1363 MergeBasicBlockIntoOnlyPred(DestBB, nullptr);
1365 if (isEntry && BB != &BB->getParent()->getEntryBlock())
1366 BB->moveBefore(&BB->getParent()->getEntryBlock());
1368 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1373 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
1374 // to handle the new incoming edges it is about to have.
1376 for (BasicBlock::iterator BBI = DestBB->begin();
1377 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1378 // Remove the incoming value for BB, and remember it.
1379 Value *InVal = PN->removeIncomingValue(BB, false);
1381 // Two options: either the InVal is a phi node defined in BB or it is some
1382 // value that dominates BB.
1383 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
1384 if (InValPhi && InValPhi->getParent() == BB) {
1385 // Add all of the input values of the input PHI as inputs of this phi.
1386 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
1387 PN->addIncoming(InValPhi->getIncomingValue(i),
1388 InValPhi->getIncomingBlock(i));
1390 // Otherwise, add one instance of the dominating value for each edge that
1391 // we will be adding.
1392 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
1393 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
1394 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
1396 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1397 PN->addIncoming(InVal, *PI);
1402 // The PHIs are now updated, change everything that refers to BB to use
1403 // DestBB and remove BB.
1404 BB->replaceAllUsesWith(DestBB);
1405 BB->eraseFromParent();
1408 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
1411 // Computes a map of base pointer relocation instructions to corresponding
1412 // derived pointer relocation instructions given a vector of all relocate calls
1413 static void computeBaseDerivedRelocateMap(
1414 const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls,
1415 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>>
1417 // Collect information in two maps: one primarily for locating the base object
1418 // while filling the second map; the second map is the final structure holding
1419 // a mapping between Base and corresponding Derived relocate calls
1420 DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap;
1421 for (auto *ThisRelocate : AllRelocateCalls) {
1422 auto K = std::make_pair(ThisRelocate->getBasePtrIndex(),
1423 ThisRelocate->getDerivedPtrIndex());
1424 RelocateIdxMap.insert(std::make_pair(K, ThisRelocate));
1426 for (auto &Item : RelocateIdxMap) {
1427 std::pair<unsigned, unsigned> Key = Item.first;
1428 if (Key.first == Key.second)
1429 // Base relocation: nothing to insert
1432 GCRelocateInst *I = Item.second;
1433 auto BaseKey = std::make_pair(Key.first, Key.first);
1435 // We're iterating over RelocateIdxMap so we cannot modify it.
1436 auto MaybeBase = RelocateIdxMap.find(BaseKey);
1437 if (MaybeBase == RelocateIdxMap.end())
1438 // TODO: We might want to insert a new base object relocate and gep off
1439 // that, if there are enough derived object relocates.
1442 RelocateInstMap[MaybeBase->second].push_back(I);
1446 // Accepts a GEP and extracts the operands into a vector provided they're all
1447 // small integer constants
1448 static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP,
1449 SmallVectorImpl<Value *> &OffsetV) {
1450 for (unsigned i = 1; i < GEP->getNumOperands(); i++) {
1451 // Only accept small constant integer operands
1452 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i));
1453 if (!Op || Op->getZExtValue() > 20)
1457 for (unsigned i = 1; i < GEP->getNumOperands(); i++)
1458 OffsetV.push_back(GEP->getOperand(i));
1462 // Takes a RelocatedBase (base pointer relocation instruction) and Targets to
1463 // replace, computes a replacement, and affects it.
1465 simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase,
1466 const SmallVectorImpl<GCRelocateInst *> &Targets) {
1467 bool MadeChange = false;
1468 for (GCRelocateInst *ToReplace : Targets) {
1469 assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&
1470 "Not relocating a derived object of the original base object");
1471 if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) {
1472 // A duplicate relocate call. TODO: coalesce duplicates.
1476 if (RelocatedBase->getParent() != ToReplace->getParent()) {
1477 // Base and derived relocates are in different basic blocks.
1478 // In this case transform is only valid when base dominates derived
1479 // relocate. However it would be too expensive to check dominance
1480 // for each such relocate, so we skip the whole transformation.
1484 Value *Base = ToReplace->getBasePtr();
1485 auto Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr());
1486 if (!Derived || Derived->getPointerOperand() != Base)
1489 SmallVector<Value *, 2> OffsetV;
1490 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV))
1493 // Create a Builder and replace the target callsite with a gep
1494 assert(RelocatedBase->getNextNode() && "Should always have one since it's not a terminator");
1496 // Insert after RelocatedBase
1497 IRBuilder<> Builder(RelocatedBase->getNextNode());
1498 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1500 // If gc_relocate does not match the actual type, cast it to the right type.
1501 // In theory, there must be a bitcast after gc_relocate if the type does not
1502 // match, and we should reuse it to get the derived pointer. But it could be
1506 // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1511 // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...)
1515 // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ]
1516 // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)*
1518 // In this case, we can not find the bitcast any more. So we insert a new bitcast
1519 // no matter there is already one or not. In this way, we can handle all cases, and
1520 // the extra bitcast should be optimized away in later passes.
1521 Value *ActualRelocatedBase = RelocatedBase;
1522 if (RelocatedBase->getType() != Base->getType()) {
1523 ActualRelocatedBase =
1524 Builder.CreateBitCast(RelocatedBase, Base->getType());
1526 Value *Replacement = Builder.CreateGEP(
1527 Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV));
1528 Replacement->takeName(ToReplace);
1529 // If the newly generated derived pointer's type does not match the original derived
1530 // pointer's type, cast the new derived pointer to match it. Same reasoning as above.
1531 Value *ActualReplacement = Replacement;
1532 if (Replacement->getType() != ToReplace->getType()) {
1534 Builder.CreateBitCast(Replacement, ToReplace->getType());
1536 ToReplace->replaceAllUsesWith(ActualReplacement);
1537 ToReplace->eraseFromParent();
1547 // %ptr = gep %base + 15
1548 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1549 // %base' = relocate(%tok, i32 4, i32 4)
1550 // %ptr' = relocate(%tok, i32 4, i32 5)
1551 // %val = load %ptr'
1556 // %ptr = gep %base + 15
1557 // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr)
1558 // %base' = gc.relocate(%tok, i32 4, i32 4)
1559 // %ptr' = gep %base' + 15
1560 // %val = load %ptr'
1561 bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) {
1562 bool MadeChange = false;
1563 SmallVector<GCRelocateInst *, 2> AllRelocateCalls;
1565 for (auto *U : I.users())
1566 if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U))
1567 // Collect all the relocate calls associated with a statepoint
1568 AllRelocateCalls.push_back(Relocate);
1570 // We need atleast one base pointer relocation + one derived pointer
1571 // relocation to mangle
1572 if (AllRelocateCalls.size() < 2)
1575 // RelocateInstMap is a mapping from the base relocate instruction to the
1576 // corresponding derived relocate instructions
1577 DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap;
1578 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap);
1579 if (RelocateInstMap.empty())
1582 for (auto &Item : RelocateInstMap)
1583 // Item.first is the RelocatedBase to offset against
1584 // Item.second is the vector of Targets to replace
1585 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second);
1589 /// SinkCast - Sink the specified cast instruction into its user blocks
1590 static bool SinkCast(CastInst *CI) {
1591 BasicBlock *DefBB = CI->getParent();
1593 /// InsertedCasts - Only insert a cast in each block once.
1594 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
1596 bool MadeChange = false;
1597 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1599 Use &TheUse = UI.getUse();
1600 Instruction *User = cast<Instruction>(*UI);
1602 // Figure out which BB this cast is used in. For PHI's this is the
1603 // appropriate predecessor block.
1604 BasicBlock *UserBB = User->getParent();
1605 if (PHINode *PN = dyn_cast<PHINode>(User)) {
1606 UserBB = PN->getIncomingBlock(TheUse);
1609 // Preincrement use iterator so we don't invalidate it.
1612 // If the block selected to receive the cast is an EH pad that does not
1613 // allow non-PHI instructions before the terminator, we can't sink the
1615 if (UserBB->getTerminator()->isEHPad())
1618 // If this user is in the same block as the cast, don't change the cast.
1619 if (UserBB == DefBB) continue;
1621 // If we have already inserted a cast into this block, use it.
1622 CastInst *&InsertedCast = InsertedCasts[UserBB];
1624 if (!InsertedCast) {
1625 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1626 assert(InsertPt != UserBB->end());
1627 InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0),
1628 CI->getType(), "", &*InsertPt);
1631 // Replace a use of the cast with a use of the new cast.
1632 TheUse = InsertedCast;
1637 // If we removed all uses, nuke the cast.
1638 if (CI->use_empty()) {
1639 CI->eraseFromParent();
1646 /// If the specified cast instruction is a noop copy (e.g. it's casting from
1647 /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to
1648 /// reduce the number of virtual registers that must be created and coalesced.
1650 /// Return true if any changes are made.
1652 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI,
1653 const DataLayout &DL) {
1654 // If this is a noop copy,
1655 EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1656 EVT DstVT = TLI.getValueType(DL, CI->getType());
1658 // This is an fp<->int conversion?
1659 if (SrcVT.isInteger() != DstVT.isInteger())
1662 // If this is an extension, it will be a zero or sign extension, which
1664 if (SrcVT.bitsLT(DstVT)) return false;
1666 // If these values will be promoted, find out what they will be promoted
1667 // to. This helps us consider truncates on PPC as noop copies when they
1669 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
1670 TargetLowering::TypePromoteInteger)
1671 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
1672 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
1673 TargetLowering::TypePromoteInteger)
1674 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
1676 // If, after promotion, these are the same types, this is a noop copy.
1680 return SinkCast(CI);
1683 /// Try to combine CI into a call to the llvm.uadd.with.overflow intrinsic if
1686 /// Return true if any changes were made.
1687 static bool CombineUAddWithOverflow(CmpInst *CI) {
1691 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI))))
1694 Type *Ty = AddI->getType();
1695 if (!isa<IntegerType>(Ty))
1698 // We don't want to move around uses of condition values this late, so we we
1699 // check if it is legal to create the call to the intrinsic in the basic
1700 // block containing the icmp:
1702 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse())
1706 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption
1708 if (AddI->hasOneUse())
1709 assert(*AddI->user_begin() == CI && "expected!");
1712 Module *M = CI->getModule();
1713 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty);
1715 auto *InsertPt = AddI->hasOneUse() ? CI : AddI;
1717 auto *UAddWithOverflow =
1718 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt);
1719 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt);
1721 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt);
1723 CI->replaceAllUsesWith(Overflow);
1724 AddI->replaceAllUsesWith(UAdd);
1725 CI->eraseFromParent();
1726 AddI->eraseFromParent();
1730 /// Sink the given CmpInst into user blocks to reduce the number of virtual
1731 /// registers that must be created and coalesced. This is a clear win except on
1732 /// targets with multiple condition code registers (PowerPC), where it might
1733 /// lose; some adjustment may be wanted there.
1735 /// Return true if any changes are made.
1736 static bool SinkCmpExpression(CmpInst *CI) {
1737 BasicBlock *DefBB = CI->getParent();
1739 /// Only insert a cmp in each block once.
1740 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
1742 bool MadeChange = false;
1743 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end();
1745 Use &TheUse = UI.getUse();
1746 Instruction *User = cast<Instruction>(*UI);
1748 // Preincrement use iterator so we don't invalidate it.
1751 // Don't bother for PHI nodes.
1752 if (isa<PHINode>(User))
1755 // Figure out which BB this cmp is used in.
1756 BasicBlock *UserBB = User->getParent();
1758 // If this user is in the same block as the cmp, don't change the cmp.
1759 if (UserBB == DefBB) continue;
1761 // If we have already inserted a cmp into this block, use it.
1762 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
1765 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1766 assert(InsertPt != UserBB->end());
1768 CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1769 CI->getOperand(0), CI->getOperand(1), "", &*InsertPt);
1772 // Replace a use of the cmp with a use of the new cmp.
1773 TheUse = InsertedCmp;
1778 // If we removed all uses, nuke the cmp.
1779 if (CI->use_empty()) {
1780 CI->eraseFromParent();
1787 static bool OptimizeCmpExpression(CmpInst *CI) {
1788 if (SinkCmpExpression(CI))
1791 if (CombineUAddWithOverflow(CI))
1797 /// Check if the candidates could be combined with a shift instruction, which
1799 /// 1. Truncate instruction
1800 /// 2. And instruction and the imm is a mask of the low bits:
1801 /// imm & (imm+1) == 0
1802 static bool isExtractBitsCandidateUse(Instruction *User) {
1803 if (!isa<TruncInst>(User)) {
1804 if (User->getOpcode() != Instruction::And ||
1805 !isa<ConstantInt>(User->getOperand(1)))
1808 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue();
1810 if ((Cimm & (Cimm + 1)).getBoolValue())
1816 /// Sink both shift and truncate instruction to the use of truncate's BB.
1818 SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI,
1819 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts,
1820 const TargetLowering &TLI, const DataLayout &DL) {
1821 BasicBlock *UserBB = User->getParent();
1822 DenseMap<BasicBlock *, CastInst *> InsertedTruncs;
1823 TruncInst *TruncI = dyn_cast<TruncInst>(User);
1824 bool MadeChange = false;
1826 for (Value::user_iterator TruncUI = TruncI->user_begin(),
1827 TruncE = TruncI->user_end();
1828 TruncUI != TruncE;) {
1830 Use &TruncTheUse = TruncUI.getUse();
1831 Instruction *TruncUser = cast<Instruction>(*TruncUI);
1832 // Preincrement use iterator so we don't invalidate it.
1836 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode());
1840 // If the use is actually a legal node, there will not be an
1841 // implicit truncate.
1842 // FIXME: always querying the result type is just an
1843 // approximation; some nodes' legality is determined by the
1844 // operand or other means. There's no good way to find out though.
1845 if (TLI.isOperationLegalOrCustom(
1846 ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true)))
1849 // Don't bother for PHI nodes.
1850 if (isa<PHINode>(TruncUser))
1853 BasicBlock *TruncUserBB = TruncUser->getParent();
1855 if (UserBB == TruncUserBB)
1858 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB];
1859 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB];
1861 if (!InsertedShift && !InsertedTrunc) {
1862 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt();
1863 assert(InsertPt != TruncUserBB->end());
1865 if (ShiftI->getOpcode() == Instruction::AShr)
1866 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1869 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1873 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt();
1875 assert(TruncInsertPt != TruncUserBB->end());
1877 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift,
1878 TruncI->getType(), "", &*TruncInsertPt);
1882 TruncTheUse = InsertedTrunc;
1888 /// Sink the shift *right* instruction into user blocks if the uses could
1889 /// potentially be combined with this shift instruction and generate BitExtract
1890 /// instruction. It will only be applied if the architecture supports BitExtract
1891 /// instruction. Here is an example:
1893 /// %x.extract.shift = lshr i64 %arg1, 32
1895 /// %x.extract.trunc = trunc i64 %x.extract.shift to i16
1899 /// %x.extract.shift.1 = lshr i64 %arg1, 32
1900 /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16
1902 /// CodeGen will recoginze the pattern in BB2 and generate BitExtract
1904 /// Return true if any changes are made.
1905 static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI,
1906 const TargetLowering &TLI,
1907 const DataLayout &DL) {
1908 BasicBlock *DefBB = ShiftI->getParent();
1910 /// Only insert instructions in each block once.
1911 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts;
1913 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType()));
1915 bool MadeChange = false;
1916 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end();
1918 Use &TheUse = UI.getUse();
1919 Instruction *User = cast<Instruction>(*UI);
1920 // Preincrement use iterator so we don't invalidate it.
1923 // Don't bother for PHI nodes.
1924 if (isa<PHINode>(User))
1927 if (!isExtractBitsCandidateUse(User))
1930 BasicBlock *UserBB = User->getParent();
1932 if (UserBB == DefBB) {
1933 // If the shift and truncate instruction are in the same BB. The use of
1934 // the truncate(TruncUse) may still introduce another truncate if not
1935 // legal. In this case, we would like to sink both shift and truncate
1936 // instruction to the BB of TruncUse.
1939 // i64 shift.result = lshr i64 opnd, imm
1940 // trunc.result = trunc shift.result to i16
1943 // ----> We will have an implicit truncate here if the architecture does
1944 // not have i16 compare.
1945 // cmp i16 trunc.result, opnd2
1947 if (isa<TruncInst>(User) && shiftIsLegal
1948 // If the type of the truncate is legal, no trucate will be
1949 // introduced in other basic blocks.
1951 (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType()))))
1953 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL);
1957 // If we have already inserted a shift into this block, use it.
1958 BinaryOperator *&InsertedShift = InsertedShifts[UserBB];
1960 if (!InsertedShift) {
1961 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1962 assert(InsertPt != UserBB->end());
1964 if (ShiftI->getOpcode() == Instruction::AShr)
1965 InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI,
1968 InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI,
1974 // Replace a use of the shift with a use of the new shift.
1975 TheUse = InsertedShift;
1978 // If we removed all uses, nuke the shift.
1979 if (ShiftI->use_empty())
1980 ShiftI->eraseFromParent();
1985 // Translate a masked load intrinsic like
1986 // <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align,
1987 // <16 x i1> %mask, <16 x i32> %passthru)
1988 // to a chain of basic blocks, with loading element one-by-one if
1989 // the appropriate mask bit is set
1991 // %1 = bitcast i8* %addr to i32*
1992 // %2 = extractelement <16 x i1> %mask, i32 0
1993 // %3 = icmp eq i1 %2, true
1994 // br i1 %3, label %cond.load, label %else
1996 //cond.load: ; preds = %0
1997 // %4 = getelementptr i32* %1, i32 0
1998 // %5 = load i32* %4
1999 // %6 = insertelement <16 x i32> undef, i32 %5, i32 0
2002 //else: ; preds = %0, %cond.load
2003 // %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ]
2004 // %7 = extractelement <16 x i1> %mask, i32 1
2005 // %8 = icmp eq i1 %7, true
2006 // br i1 %8, label %cond.load1, label %else2
2008 //cond.load1: ; preds = %else
2009 // %9 = getelementptr i32* %1, i32 1
2010 // %10 = load i32* %9
2011 // %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1
2014 //else2: ; preds = %else, %cond.load1
2015 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2016 // %12 = extractelement <16 x i1> %mask, i32 2
2017 // %13 = icmp eq i1 %12, true
2018 // br i1 %13, label %cond.load4, label %else5
2020 static void ScalarizeMaskedLoad(CallInst *CI) {
2021 Value *Ptr = CI->getArgOperand(0);
2022 Value *Alignment = CI->getArgOperand(1);
2023 Value *Mask = CI->getArgOperand(2);
2024 Value *Src0 = CI->getArgOperand(3);
2026 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2027 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2028 assert(VecType && "Unexpected return type of masked load intrinsic");
2030 Type *EltTy = CI->getType()->getVectorElementType();
2032 IRBuilder<> Builder(CI->getContext());
2033 Instruction *InsertPt = CI;
2034 BasicBlock *IfBlock = CI->getParent();
2035 BasicBlock *CondBlock = nullptr;
2036 BasicBlock *PrevIfBlock = CI->getParent();
2038 Builder.SetInsertPoint(InsertPt);
2039 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2041 // Short-cut if the mask is all-true.
2042 bool IsAllOnesMask = isa<Constant>(Mask) &&
2043 cast<Constant>(Mask)->isAllOnesValue();
2045 if (IsAllOnesMask) {
2046 Value *NewI = Builder.CreateAlignedLoad(Ptr, AlignVal);
2047 CI->replaceAllUsesWith(NewI);
2048 CI->eraseFromParent();
2052 // Adjust alignment for the scalar instruction.
2053 AlignVal = std::min(AlignVal, VecType->getScalarSizeInBits()/8);
2054 // Bitcast %addr fron i8* to EltTy*
2056 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2057 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2058 unsigned VectorWidth = VecType->getNumElements();
2060 Value *UndefVal = UndefValue::get(VecType);
2062 // The result vector
2063 Value *VResult = UndefVal;
2065 if (isa<ConstantVector>(Mask)) {
2066 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2067 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2070 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2071 LoadInst* Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2072 VResult = Builder.CreateInsertElement(VResult, Load,
2073 Builder.getInt32(Idx));
2075 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2076 CI->replaceAllUsesWith(NewI);
2077 CI->eraseFromParent();
2081 PHINode *Phi = nullptr;
2082 Value *PrevPhi = UndefVal;
2084 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2086 // Fill the "else" block, created in the previous iteration
2088 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ]
2089 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2090 // %to_load = icmp eq i1 %mask_1, true
2091 // br i1 %to_load, label %cond.load, label %else
2094 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2095 Phi->addIncoming(VResult, CondBlock);
2096 Phi->addIncoming(PrevPhi, PrevIfBlock);
2101 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2102 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2103 ConstantInt::get(Predicate->getType(), 1));
2105 // Create "cond" block
2107 // %EltAddr = getelementptr i32* %1, i32 0
2108 // %Elt = load i32* %EltAddr
2109 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2111 CondBlock = IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.load");
2112 Builder.SetInsertPoint(InsertPt);
2115 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2116 LoadInst *Load = Builder.CreateAlignedLoad(Gep, AlignVal);
2117 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx));
2119 // Create "else" block, fill it in the next iteration
2120 BasicBlock *NewIfBlock =
2121 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2122 Builder.SetInsertPoint(InsertPt);
2123 Instruction *OldBr = IfBlock->getTerminator();
2124 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2125 OldBr->eraseFromParent();
2126 PrevIfBlock = IfBlock;
2127 IfBlock = NewIfBlock;
2130 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2131 Phi->addIncoming(VResult, CondBlock);
2132 Phi->addIncoming(PrevPhi, PrevIfBlock);
2133 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2134 CI->replaceAllUsesWith(NewI);
2135 CI->eraseFromParent();
2138 // Translate a masked store intrinsic, like
2139 // void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align,
2141 // to a chain of basic blocks, that stores element one-by-one if
2142 // the appropriate mask bit is set
2144 // %1 = bitcast i8* %addr to i32*
2145 // %2 = extractelement <16 x i1> %mask, i32 0
2146 // %3 = icmp eq i1 %2, true
2147 // br i1 %3, label %cond.store, label %else
2149 // cond.store: ; preds = %0
2150 // %4 = extractelement <16 x i32> %val, i32 0
2151 // %5 = getelementptr i32* %1, i32 0
2152 // store i32 %4, i32* %5
2155 // else: ; preds = %0, %cond.store
2156 // %6 = extractelement <16 x i1> %mask, i32 1
2157 // %7 = icmp eq i1 %6, true
2158 // br i1 %7, label %cond.store1, label %else2
2160 // cond.store1: ; preds = %else
2161 // %8 = extractelement <16 x i32> %val, i32 1
2162 // %9 = getelementptr i32* %1, i32 1
2163 // store i32 %8, i32* %9
2166 static void ScalarizeMaskedStore(CallInst *CI) {
2167 Value *Src = CI->getArgOperand(0);
2168 Value *Ptr = CI->getArgOperand(1);
2169 Value *Alignment = CI->getArgOperand(2);
2170 Value *Mask = CI->getArgOperand(3);
2172 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2173 VectorType *VecType = dyn_cast<VectorType>(Src->getType());
2174 assert(VecType && "Unexpected data type in masked store intrinsic");
2176 Type *EltTy = VecType->getElementType();
2178 IRBuilder<> Builder(CI->getContext());
2179 Instruction *InsertPt = CI;
2180 BasicBlock *IfBlock = CI->getParent();
2181 Builder.SetInsertPoint(InsertPt);
2182 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2184 // Short-cut if the mask is all-true.
2185 bool IsAllOnesMask = isa<Constant>(Mask) &&
2186 cast<Constant>(Mask)->isAllOnesValue();
2188 if (IsAllOnesMask) {
2189 Builder.CreateAlignedStore(Src, Ptr, AlignVal);
2190 CI->eraseFromParent();
2194 // Adjust alignment for the scalar instruction.
2195 AlignVal = std::max(AlignVal, VecType->getScalarSizeInBits()/8);
2196 // Bitcast %addr fron i8* to EltTy*
2198 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace());
2199 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType);
2200 unsigned VectorWidth = VecType->getNumElements();
2202 if (isa<ConstantVector>(Mask)) {
2203 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2204 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2206 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2208 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2209 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2211 CI->eraseFromParent();
2215 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2217 // Fill the "else" block, created in the previous iteration
2219 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx
2220 // %to_store = icmp eq i1 %mask_1, true
2221 // br i1 %to_store, label %cond.store, label %else
2223 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx));
2224 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2225 ConstantInt::get(Predicate->getType(), 1));
2227 // Create "cond" block
2229 // %OneElt = extractelement <16 x i32> %Src, i32 Idx
2230 // %EltAddr = getelementptr i32* %1, i32 0
2231 // %store i32 %OneElt, i32* %EltAddr
2233 BasicBlock *CondBlock =
2234 IfBlock->splitBasicBlock(InsertPt->getIterator(), "cond.store");
2235 Builder.SetInsertPoint(InsertPt);
2237 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx));
2239 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx));
2240 Builder.CreateAlignedStore(OneElt, Gep, AlignVal);
2242 // Create "else" block, fill it in the next iteration
2243 BasicBlock *NewIfBlock =
2244 CondBlock->splitBasicBlock(InsertPt->getIterator(), "else");
2245 Builder.SetInsertPoint(InsertPt);
2246 Instruction *OldBr = IfBlock->getTerminator();
2247 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2248 OldBr->eraseFromParent();
2249 IfBlock = NewIfBlock;
2251 CI->eraseFromParent();
2254 // Translate a masked gather intrinsic like
2255 // <16 x i32 > @llvm.masked.gather.v16i32( <16 x i32*> %Ptrs, i32 4,
2256 // <16 x i1> %Mask, <16 x i32> %Src)
2257 // to a chain of basic blocks, with loading element one-by-one if
2258 // the appropriate mask bit is set
2260 // % Ptrs = getelementptr i32, i32* %base, <16 x i64> %ind
2261 // % Mask0 = extractelement <16 x i1> %Mask, i32 0
2262 // % ToLoad0 = icmp eq i1 % Mask0, true
2263 // br i1 % ToLoad0, label %cond.load, label %else
2266 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2267 // % Load0 = load i32, i32* % Ptr0, align 4
2268 // % Res0 = insertelement <16 x i32> undef, i32 % Load0, i32 0
2272 // %res.phi.else = phi <16 x i32>[% Res0, %cond.load], [undef, % 0]
2273 // % Mask1 = extractelement <16 x i1> %Mask, i32 1
2274 // % ToLoad1 = icmp eq i1 % Mask1, true
2275 // br i1 % ToLoad1, label %cond.load1, label %else2
2278 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2279 // % Load1 = load i32, i32* % Ptr1, align 4
2280 // % Res1 = insertelement <16 x i32> %res.phi.else, i32 % Load1, i32 1
2283 // % Result = select <16 x i1> %Mask, <16 x i32> %res.phi.select, <16 x i32> %Src
2284 // ret <16 x i32> %Result
2285 static void ScalarizeMaskedGather(CallInst *CI) {
2286 Value *Ptrs = CI->getArgOperand(0);
2287 Value *Alignment = CI->getArgOperand(1);
2288 Value *Mask = CI->getArgOperand(2);
2289 Value *Src0 = CI->getArgOperand(3);
2291 VectorType *VecType = dyn_cast<VectorType>(CI->getType());
2293 assert(VecType && "Unexpected return type of masked load intrinsic");
2295 IRBuilder<> Builder(CI->getContext());
2296 Instruction *InsertPt = CI;
2297 BasicBlock *IfBlock = CI->getParent();
2298 BasicBlock *CondBlock = nullptr;
2299 BasicBlock *PrevIfBlock = CI->getParent();
2300 Builder.SetInsertPoint(InsertPt);
2301 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2303 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2305 Value *UndefVal = UndefValue::get(VecType);
2307 // The result vector
2308 Value *VResult = UndefVal;
2309 unsigned VectorWidth = VecType->getNumElements();
2311 // Shorten the way if the mask is a vector of constants.
2312 bool IsConstMask = isa<ConstantVector>(Mask);
2315 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2316 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2318 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2319 "Ptr" + Twine(Idx));
2320 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2321 "Load" + Twine(Idx));
2322 VResult = Builder.CreateInsertElement(VResult, Load,
2323 Builder.getInt32(Idx),
2324 "Res" + Twine(Idx));
2326 Value *NewI = Builder.CreateSelect(Mask, VResult, Src0);
2327 CI->replaceAllUsesWith(NewI);
2328 CI->eraseFromParent();
2332 PHINode *Phi = nullptr;
2333 Value *PrevPhi = UndefVal;
2335 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2337 // Fill the "else" block, created in the previous iteration
2339 // %Mask1 = extractelement <16 x i1> %Mask, i32 1
2340 // %ToLoad1 = icmp eq i1 %Mask1, true
2341 // br i1 %ToLoad1, label %cond.load, label %else
2344 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else");
2345 Phi->addIncoming(VResult, CondBlock);
2346 Phi->addIncoming(PrevPhi, PrevIfBlock);
2351 Value *Predicate = Builder.CreateExtractElement(Mask,
2352 Builder.getInt32(Idx),
2353 "Mask" + Twine(Idx));
2354 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2355 ConstantInt::get(Predicate->getType(), 1),
2356 "ToLoad" + Twine(Idx));
2358 // Create "cond" block
2360 // %EltAddr = getelementptr i32* %1, i32 0
2361 // %Elt = load i32* %EltAddr
2362 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx
2364 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load");
2365 Builder.SetInsertPoint(InsertPt);
2367 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2368 "Ptr" + Twine(Idx));
2369 LoadInst *Load = Builder.CreateAlignedLoad(Ptr, AlignVal,
2370 "Load" + Twine(Idx));
2371 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx),
2372 "Res" + Twine(Idx));
2374 // Create "else" block, fill it in the next iteration
2375 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2376 Builder.SetInsertPoint(InsertPt);
2377 Instruction *OldBr = IfBlock->getTerminator();
2378 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2379 OldBr->eraseFromParent();
2380 PrevIfBlock = IfBlock;
2381 IfBlock = NewIfBlock;
2384 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select");
2385 Phi->addIncoming(VResult, CondBlock);
2386 Phi->addIncoming(PrevPhi, PrevIfBlock);
2387 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0);
2388 CI->replaceAllUsesWith(NewI);
2389 CI->eraseFromParent();
2392 // Translate a masked scatter intrinsic, like
2393 // void @llvm.masked.scatter.v16i32(<16 x i32> %Src, <16 x i32*>* %Ptrs, i32 4,
2395 // to a chain of basic blocks, that stores element one-by-one if
2396 // the appropriate mask bit is set.
2398 // % Ptrs = getelementptr i32, i32* %ptr, <16 x i64> %ind
2399 // % Mask0 = extractelement <16 x i1> % Mask, i32 0
2400 // % ToStore0 = icmp eq i1 % Mask0, true
2401 // br i1 %ToStore0, label %cond.store, label %else
2404 // % Elt0 = extractelement <16 x i32> %Src, i32 0
2405 // % Ptr0 = extractelement <16 x i32*> %Ptrs, i32 0
2406 // store i32 %Elt0, i32* % Ptr0, align 4
2410 // % Mask1 = extractelement <16 x i1> % Mask, i32 1
2411 // % ToStore1 = icmp eq i1 % Mask1, true
2412 // br i1 % ToStore1, label %cond.store1, label %else2
2415 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2416 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2417 // store i32 % Elt1, i32* % Ptr1, align 4
2420 static void ScalarizeMaskedScatter(CallInst *CI) {
2421 Value *Src = CI->getArgOperand(0);
2422 Value *Ptrs = CI->getArgOperand(1);
2423 Value *Alignment = CI->getArgOperand(2);
2424 Value *Mask = CI->getArgOperand(3);
2426 assert(isa<VectorType>(Src->getType()) &&
2427 "Unexpected data type in masked scatter intrinsic");
2428 assert(isa<VectorType>(Ptrs->getType()) &&
2429 isa<PointerType>(Ptrs->getType()->getVectorElementType()) &&
2430 "Vector of pointers is expected in masked scatter intrinsic");
2432 IRBuilder<> Builder(CI->getContext());
2433 Instruction *InsertPt = CI;
2434 BasicBlock *IfBlock = CI->getParent();
2435 Builder.SetInsertPoint(InsertPt);
2436 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
2438 unsigned AlignVal = cast<ConstantInt>(Alignment)->getZExtValue();
2439 unsigned VectorWidth = Src->getType()->getVectorNumElements();
2441 // Shorten the way if the mask is a vector of constants.
2442 bool IsConstMask = isa<ConstantVector>(Mask);
2445 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2446 if (cast<ConstantVector>(Mask)->getOperand(Idx)->isNullValue())
2448 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2449 "Elt" + Twine(Idx));
2450 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2451 "Ptr" + Twine(Idx));
2452 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2454 CI->eraseFromParent();
2457 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) {
2458 // Fill the "else" block, created in the previous iteration
2460 // % Mask1 = extractelement <16 x i1> % Mask, i32 Idx
2461 // % ToStore = icmp eq i1 % Mask1, true
2462 // br i1 % ToStore, label %cond.store, label %else
2464 Value *Predicate = Builder.CreateExtractElement(Mask,
2465 Builder.getInt32(Idx),
2466 "Mask" + Twine(Idx));
2468 Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate,
2469 ConstantInt::get(Predicate->getType(), 1),
2470 "ToStore" + Twine(Idx));
2472 // Create "cond" block
2474 // % Elt1 = extractelement <16 x i32> %Src, i32 1
2475 // % Ptr1 = extractelement <16 x i32*> %Ptrs, i32 1
2476 // %store i32 % Elt1, i32* % Ptr1
2478 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
2479 Builder.SetInsertPoint(InsertPt);
2481 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx),
2482 "Elt" + Twine(Idx));
2483 Value *Ptr = Builder.CreateExtractElement(Ptrs, Builder.getInt32(Idx),
2484 "Ptr" + Twine(Idx));
2485 Builder.CreateAlignedStore(OneElt, Ptr, AlignVal);
2487 // Create "else" block, fill it in the next iteration
2488 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
2489 Builder.SetInsertPoint(InsertPt);
2490 Instruction *OldBr = IfBlock->getTerminator();
2491 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
2492 OldBr->eraseFromParent();
2493 IfBlock = NewIfBlock;
2495 CI->eraseFromParent();
2498 /// If counting leading or trailing zeros is an expensive operation and a zero
2499 /// input is defined, add a check for zero to avoid calling the intrinsic.
2501 /// We want to transform:
2502 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false)
2506 /// %cmpz = icmp eq i64 %A, 0
2507 /// br i1 %cmpz, label %cond.end, label %cond.false
2509 /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true)
2510 /// br label %cond.end
2512 /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ]
2514 /// If the transform is performed, return true and set ModifiedDT to true.
2515 static bool despeculateCountZeros(IntrinsicInst *CountZeros,
2516 const TargetLowering *TLI,
2517 const DataLayout *DL,
2522 // If a zero input is undefined, it doesn't make sense to despeculate that.
2523 if (match(CountZeros->getOperand(1), m_One()))
2526 // If it's cheap to speculate, there's nothing to do.
2527 auto IntrinsicID = CountZeros->getIntrinsicID();
2528 if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) ||
2529 (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz()))
2532 // Only handle legal scalar cases. Anything else requires too much work.
2533 Type *Ty = CountZeros->getType();
2534 unsigned SizeInBits = Ty->getPrimitiveSizeInBits();
2535 if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSize())
2538 // The intrinsic will be sunk behind a compare against zero and branch.
2539 BasicBlock *StartBlock = CountZeros->getParent();
2540 BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");
2542 // Create another block after the count zero intrinsic. A PHI will be added
2543 // in this block to select the result of the intrinsic or the bit-width
2544 // constant if the input to the intrinsic is zero.
2545 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros));
2546 BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end");
2548 // Set up a builder to create a compare, conditional branch, and PHI.
2549 IRBuilder<> Builder(CountZeros->getContext());
2550 Builder.SetInsertPoint(StartBlock->getTerminator());
2551 Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc());
2553 // Replace the unconditional branch that was created by the first split with
2554 // a compare against zero and a conditional branch.
2555 Value *Zero = Constant::getNullValue(Ty);
2556 Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz");
2557 Builder.CreateCondBr(Cmp, EndBlock, CallBlock);
2558 StartBlock->getTerminator()->eraseFromParent();
2560 // Create a PHI in the end block to select either the output of the intrinsic
2561 // or the bit width of the operand.
2562 Builder.SetInsertPoint(&EndBlock->front());
2563 PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz");
2564 CountZeros->replaceAllUsesWith(PN);
2565 Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits));
2566 PN->addIncoming(BitWidth, StartBlock);
2567 PN->addIncoming(CountZeros, CallBlock);
2569 // We are explicitly handling the zero case, so we can set the intrinsic's
2570 // undefined zero argument to 'true'. This will also prevent reprocessing the
2571 // intrinsic; we only despeculate when a zero input is defined.
2572 CountZeros->setArgOperand(1, Builder.getTrue());
2577 bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool& ModifiedDT) {
2578 BasicBlock *BB = CI->getParent();
2580 // Lower inline assembly if we can.
2581 // If we found an inline asm expession, and if the target knows how to
2582 // lower it to normal LLVM code, do so now.
2583 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
2584 if (TLI->ExpandInlineAsm(CI)) {
2585 // Avoid invalidating the iterator.
2586 CurInstIterator = BB->begin();
2587 // Avoid processing instructions out of order, which could cause
2588 // reuse before a value is defined.
2592 // Sink address computing for memory operands into the block.
2593 if (optimizeInlineAsmInst(CI))
2597 // Align the pointer arguments to this call if the target thinks it's a good
2599 unsigned MinSize, PrefAlign;
2600 if (TLI && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
2601 for (auto &Arg : CI->arg_operands()) {
2602 // We want to align both objects whose address is used directly and
2603 // objects whose address is used in casts and GEPs, though it only makes
2604 // sense for GEPs if the offset is a multiple of the desired alignment and
2605 // if size - offset meets the size threshold.
2606 if (!Arg->getType()->isPointerTy())
2608 APInt Offset(DL->getPointerSizeInBits(
2609 cast<PointerType>(Arg->getType())->getAddressSpace()),
2611 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset);
2612 uint64_t Offset2 = Offset.getLimitedValue();
2613 if ((Offset2 & (PrefAlign-1)) != 0)
2616 if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign &&
2617 DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
2618 AI->setAlignment(PrefAlign);
2619 // Global variables can only be aligned if they are defined in this
2620 // object (i.e. they are uniquely initialized in this object), and
2621 // over-aligning global variables that have an explicit section is
2624 if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() &&
2625 GV->getAlignment() < PrefAlign &&
2626 DL->getTypeAllocSize(GV->getType()->getElementType()) >=
2628 GV->setAlignment(PrefAlign);
2630 // If this is a memcpy (or similar) then we may be able to improve the
2632 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
2633 unsigned Align = getKnownAlignment(MI->getDest(), *DL);
2634 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
2635 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *DL));
2636 if (Align > MI->getAlignment())
2637 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
2641 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
2643 switch (II->getIntrinsicID()) {
2645 case Intrinsic::objectsize: {
2646 // Lower all uses of llvm.objectsize.*
2647 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
2648 Type *ReturnTy = CI->getType();
2649 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
2651 // Substituting this can cause recursive simplifications, which can
2652 // invalidate our iterator. Use a WeakVH to hold onto it in case this
2654 WeakVH IterHandle(&*CurInstIterator);
2656 replaceAndRecursivelySimplify(CI, RetVal,
2659 // If the iterator instruction was recursively deleted, start over at the
2660 // start of the block.
2661 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
2662 CurInstIterator = BB->begin();
2667 case Intrinsic::masked_load: {
2668 // Scalarize unsupported vector masked load
2669 if (!TTI->isLegalMaskedLoad(CI->getType())) {
2670 ScalarizeMaskedLoad(CI);
2676 case Intrinsic::masked_store: {
2677 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType())) {
2678 ScalarizeMaskedStore(CI);
2684 case Intrinsic::masked_gather: {
2685 if (!TTI->isLegalMaskedGather(CI->getType())) {
2686 ScalarizeMaskedGather(CI);
2692 case Intrinsic::masked_scatter: {
2693 if (!TTI->isLegalMaskedScatter(CI->getArgOperand(0)->getType())) {
2694 ScalarizeMaskedScatter(CI);
2700 case Intrinsic::aarch64_stlxr:
2701 case Intrinsic::aarch64_stxr: {
2702 ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0));
2703 if (!ExtVal || !ExtVal->hasOneUse() ||
2704 ExtVal->getParent() == CI->getParent())
2706 // Sink a zext feeding stlxr/stxr before it, so it can be folded into it.
2707 ExtVal->moveBefore(CI);
2708 // Mark this instruction as "inserted by CGP", so that other
2709 // optimizations don't touch it.
2710 InsertedInsts.insert(ExtVal);
2713 case Intrinsic::invariant_group_barrier:
2714 II->replaceAllUsesWith(II->getArgOperand(0));
2715 II->eraseFromParent();
2718 case Intrinsic::cttz:
2719 case Intrinsic::ctlz:
2720 // If counting zeros is expensive, try to avoid it.
2721 return despeculateCountZeros(II, TLI, DL, ModifiedDT);
2725 // Unknown address space.
2726 // TODO: Target hook to pick which address space the intrinsic cares
2728 unsigned AddrSpace = ~0u;
2729 SmallVector<Value*, 2> PtrOps;
2731 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy, AddrSpace))
2732 while (!PtrOps.empty())
2733 if (optimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy, AddrSpace))
2738 // From here on out we're working with named functions.
2739 if (!CI->getCalledFunction()) return false;
2741 // Lower all default uses of _chk calls. This is very similar
2742 // to what InstCombineCalls does, but here we are only lowering calls
2743 // to fortified library functions (e.g. __memcpy_chk) that have the default
2744 // "don't know" as the objectsize. Anything else should be left alone.
2745 FortifiedLibCallSimplifier Simplifier(TLInfo, true);
2746 if (Value *V = Simplifier.optimizeCall(CI)) {
2747 CI->replaceAllUsesWith(V);
2748 CI->eraseFromParent();
2754 /// Look for opportunities to duplicate return instructions to the predecessor
2755 /// to enable tail call optimizations. The case it is currently looking for is:
2758 /// %tmp0 = tail call i32 @f0()
2759 /// br label %return
2761 /// %tmp1 = tail call i32 @f1()
2762 /// br label %return
2764 /// %tmp2 = tail call i32 @f2()
2765 /// br label %return
2767 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
2775 /// %tmp0 = tail call i32 @f0()
2778 /// %tmp1 = tail call i32 @f1()
2781 /// %tmp2 = tail call i32 @f2()
2784 bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB) {
2788 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
2792 PHINode *PN = nullptr;
2793 BitCastInst *BCI = nullptr;
2794 Value *V = RI->getReturnValue();
2796 BCI = dyn_cast<BitCastInst>(V);
2798 V = BCI->getOperand(0);
2800 PN = dyn_cast<PHINode>(V);
2805 if (PN && PN->getParent() != BB)
2808 // It's not safe to eliminate the sign / zero extension of the return value.
2809 // See llvm::isInTailCallPosition().
2810 const Function *F = BB->getParent();
2811 AttributeSet CallerAttrs = F->getAttributes();
2812 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
2813 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
2816 // Make sure there are no instructions between the PHI and return, or that the
2817 // return is the first instruction in the block.
2819 BasicBlock::iterator BI = BB->begin();
2820 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
2822 // Also skip over the bitcast.
2827 BasicBlock::iterator BI = BB->begin();
2828 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
2833 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
2835 SmallVector<CallInst*, 4> TailCalls;
2837 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
2838 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
2839 // Make sure the phi value is indeed produced by the tail call.
2840 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
2841 TLI->mayBeEmittedAsTailCall(CI))
2842 TailCalls.push_back(CI);
2845 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
2846 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
2847 if (!VisitedBBs.insert(*PI).second)
2850 BasicBlock::InstListType &InstList = (*PI)->getInstList();
2851 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
2852 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
2853 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
2857 CallInst *CI = dyn_cast<CallInst>(&*RI);
2858 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
2859 TailCalls.push_back(CI);
2863 bool Changed = false;
2864 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
2865 CallInst *CI = TailCalls[i];
2868 // Conservatively require the attributes of the call to match those of the
2869 // return. Ignore noalias because it doesn't affect the call sequence.
2870 AttributeSet CalleeAttrs = CS.getAttributes();
2871 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
2872 removeAttribute(Attribute::NoAlias) !=
2873 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
2874 removeAttribute(Attribute::NoAlias))
2877 // Make sure the call instruction is followed by an unconditional branch to
2878 // the return block.
2879 BasicBlock *CallBB = CI->getParent();
2880 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
2881 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
2884 // Duplicate the return into CallBB.
2885 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
2886 ModifiedDT = Changed = true;
2890 // If we eliminated all predecessors of the block, delete the block now.
2891 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
2892 BB->eraseFromParent();
2897 //===----------------------------------------------------------------------===//
2898 // Memory Optimization
2899 //===----------------------------------------------------------------------===//
2903 /// This is an extended version of TargetLowering::AddrMode
2904 /// which holds actual Value*'s for register values.
2905 struct ExtAddrMode : public TargetLowering::AddrMode {
2908 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {}
2909 void print(raw_ostream &OS) const;
2912 bool operator==(const ExtAddrMode& O) const {
2913 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
2914 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
2915 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
2920 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
2926 void ExtAddrMode::print(raw_ostream &OS) const {
2927 bool NeedPlus = false;
2930 OS << (NeedPlus ? " + " : "")
2932 BaseGV->printAsOperand(OS, /*PrintType=*/false);
2937 OS << (NeedPlus ? " + " : "")
2943 OS << (NeedPlus ? " + " : "")
2945 BaseReg->printAsOperand(OS, /*PrintType=*/false);
2949 OS << (NeedPlus ? " + " : "")
2951 ScaledReg->printAsOperand(OS, /*PrintType=*/false);
2957 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2958 void ExtAddrMode::dump() const {
2964 /// \brief This class provides transaction based operation on the IR.
2965 /// Every change made through this class is recorded in the internal state and
2966 /// can be undone (rollback) until commit is called.
2967 class TypePromotionTransaction {
2969 /// \brief This represents the common interface of the individual transaction.
2970 /// Each class implements the logic for doing one specific modification on
2971 /// the IR via the TypePromotionTransaction.
2972 class TypePromotionAction {
2974 /// The Instruction modified.
2978 /// \brief Constructor of the action.
2979 /// The constructor performs the related action on the IR.
2980 TypePromotionAction(Instruction *Inst) : Inst(Inst) {}
2982 virtual ~TypePromotionAction() {}
2984 /// \brief Undo the modification done by this action.
2985 /// When this method is called, the IR must be in the same state as it was
2986 /// before this action was applied.
2987 /// \pre Undoing the action works if and only if the IR is in the exact same
2988 /// state as it was directly after this action was applied.
2989 virtual void undo() = 0;
2991 /// \brief Advocate every change made by this action.
2992 /// When the results on the IR of the action are to be kept, it is important
2993 /// to call this function, otherwise hidden information may be kept forever.
2994 virtual void commit() {
2995 // Nothing to be done, this action is not doing anything.
2999 /// \brief Utility to remember the position of an instruction.
3000 class InsertionHandler {
3001 /// Position of an instruction.
3002 /// Either an instruction:
3003 /// - Is the first in a basic block: BB is used.
3004 /// - Has a previous instructon: PrevInst is used.
3006 Instruction *PrevInst;
3009 /// Remember whether or not the instruction had a previous instruction.
3010 bool HasPrevInstruction;
3013 /// \brief Record the position of \p Inst.
3014 InsertionHandler(Instruction *Inst) {
3015 BasicBlock::iterator It = Inst->getIterator();
3016 HasPrevInstruction = (It != (Inst->getParent()->begin()));
3017 if (HasPrevInstruction)
3018 Point.PrevInst = &*--It;
3020 Point.BB = Inst->getParent();
3023 /// \brief Insert \p Inst at the recorded position.
3024 void insert(Instruction *Inst) {
3025 if (HasPrevInstruction) {
3026 if (Inst->getParent())
3027 Inst->removeFromParent();
3028 Inst->insertAfter(Point.PrevInst);
3030 Instruction *Position = &*Point.BB->getFirstInsertionPt();
3031 if (Inst->getParent())
3032 Inst->moveBefore(Position);
3034 Inst->insertBefore(Position);
3039 /// \brief Move an instruction before another.
3040 class InstructionMoveBefore : public TypePromotionAction {
3041 /// Original position of the instruction.
3042 InsertionHandler Position;
3045 /// \brief Move \p Inst before \p Before.
3046 InstructionMoveBefore(Instruction *Inst, Instruction *Before)
3047 : TypePromotionAction(Inst), Position(Inst) {
3048 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n");
3049 Inst->moveBefore(Before);
3052 /// \brief Move the instruction back to its original position.
3053 void undo() override {
3054 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n");
3055 Position.insert(Inst);
3059 /// \brief Set the operand of an instruction with a new value.
3060 class OperandSetter : public TypePromotionAction {
3061 /// Original operand of the instruction.
3063 /// Index of the modified instruction.
3067 /// \brief Set \p Idx operand of \p Inst with \p NewVal.
3068 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal)
3069 : TypePromotionAction(Inst), Idx(Idx) {
3070 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"
3071 << "for:" << *Inst << "\n"
3072 << "with:" << *NewVal << "\n");
3073 Origin = Inst->getOperand(Idx);
3074 Inst->setOperand(Idx, NewVal);
3077 /// \brief Restore the original value of the instruction.
3078 void undo() override {
3079 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"
3080 << "for: " << *Inst << "\n"
3081 << "with: " << *Origin << "\n");
3082 Inst->setOperand(Idx, Origin);
3086 /// \brief Hide the operands of an instruction.
3087 /// Do as if this instruction was not using any of its operands.
3088 class OperandsHider : public TypePromotionAction {
3089 /// The list of original operands.
3090 SmallVector<Value *, 4> OriginalValues;
3093 /// \brief Remove \p Inst from the uses of the operands of \p Inst.
3094 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) {
3095 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n");
3096 unsigned NumOpnds = Inst->getNumOperands();
3097 OriginalValues.reserve(NumOpnds);
3098 for (unsigned It = 0; It < NumOpnds; ++It) {
3099 // Save the current operand.
3100 Value *Val = Inst->getOperand(It);
3101 OriginalValues.push_back(Val);
3103 // We could use OperandSetter here, but that would imply an overhead
3104 // that we are not willing to pay.
3105 Inst->setOperand(It, UndefValue::get(Val->getType()));
3109 /// \brief Restore the original list of uses.
3110 void undo() override {
3111 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n");
3112 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It)
3113 Inst->setOperand(It, OriginalValues[It]);
3117 /// \brief Build a truncate instruction.
3118 class TruncBuilder : public TypePromotionAction {
3121 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty
3123 /// trunc Opnd to Ty.
3124 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) {
3125 IRBuilder<> Builder(Opnd);
3126 Val = Builder.CreateTrunc(Opnd, Ty, "promoted");
3127 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n");
3130 /// \brief Get the built value.
3131 Value *getBuiltValue() { return Val; }
3133 /// \brief Remove the built instruction.
3134 void undo() override {
3135 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n");
3136 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3137 IVal->eraseFromParent();
3141 /// \brief Build a sign extension instruction.
3142 class SExtBuilder : public TypePromotionAction {
3145 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty
3147 /// sext Opnd to Ty.
3148 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3149 : TypePromotionAction(InsertPt) {
3150 IRBuilder<> Builder(InsertPt);
3151 Val = Builder.CreateSExt(Opnd, Ty, "promoted");
3152 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n");
3155 /// \brief Get the built value.
3156 Value *getBuiltValue() { return Val; }
3158 /// \brief Remove the built instruction.
3159 void undo() override {
3160 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n");
3161 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3162 IVal->eraseFromParent();
3166 /// \brief Build a zero extension instruction.
3167 class ZExtBuilder : public TypePromotionAction {
3170 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty
3172 /// zext Opnd to Ty.
3173 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty)
3174 : TypePromotionAction(InsertPt) {
3175 IRBuilder<> Builder(InsertPt);
3176 Val = Builder.CreateZExt(Opnd, Ty, "promoted");
3177 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n");
3180 /// \brief Get the built value.
3181 Value *getBuiltValue() { return Val; }
3183 /// \brief Remove the built instruction.
3184 void undo() override {
3185 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n");
3186 if (Instruction *IVal = dyn_cast<Instruction>(Val))
3187 IVal->eraseFromParent();
3191 /// \brief Mutate an instruction to another type.
3192 class TypeMutator : public TypePromotionAction {
3193 /// Record the original type.
3197 /// \brief Mutate the type of \p Inst into \p NewTy.
3198 TypeMutator(Instruction *Inst, Type *NewTy)
3199 : TypePromotionAction(Inst), OrigTy(Inst->getType()) {
3200 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy
3202 Inst->mutateType(NewTy);
3205 /// \brief Mutate the instruction back to its original type.
3206 void undo() override {
3207 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy
3209 Inst->mutateType(OrigTy);
3213 /// \brief Replace the uses of an instruction by another instruction.
3214 class UsesReplacer : public TypePromotionAction {
3215 /// Helper structure to keep track of the replaced uses.
3216 struct InstructionAndIdx {
3217 /// The instruction using the instruction.
3219 /// The index where this instruction is used for Inst.
3221 InstructionAndIdx(Instruction *Inst, unsigned Idx)
3222 : Inst(Inst), Idx(Idx) {}
3225 /// Keep track of the original uses (pair Instruction, Index).
3226 SmallVector<InstructionAndIdx, 4> OriginalUses;
3227 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator;
3230 /// \brief Replace all the use of \p Inst by \p New.
3231 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) {
3232 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New
3234 // Record the original uses.
3235 for (Use &U : Inst->uses()) {
3236 Instruction *UserI = cast<Instruction>(U.getUser());
3237 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo()));
3239 // Now, we can replace the uses.
3240 Inst->replaceAllUsesWith(New);
3243 /// \brief Reassign the original uses of Inst to Inst.
3244 void undo() override {
3245 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n");
3246 for (use_iterator UseIt = OriginalUses.begin(),
3247 EndIt = OriginalUses.end();
3248 UseIt != EndIt; ++UseIt) {
3249 UseIt->Inst->setOperand(UseIt->Idx, Inst);
3254 /// \brief Remove an instruction from the IR.
3255 class InstructionRemover : public TypePromotionAction {
3256 /// Original position of the instruction.
3257 InsertionHandler Inserter;
3258 /// Helper structure to hide all the link to the instruction. In other
3259 /// words, this helps to do as if the instruction was removed.
3260 OperandsHider Hider;
3261 /// Keep track of the uses replaced, if any.
3262 UsesReplacer *Replacer;
3265 /// \brief Remove all reference of \p Inst and optinally replace all its
3267 /// \pre If !Inst->use_empty(), then New != nullptr
3268 InstructionRemover(Instruction *Inst, Value *New = nullptr)
3269 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst),
3272 Replacer = new UsesReplacer(Inst, New);
3273 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n");
3274 Inst->removeFromParent();
3277 ~InstructionRemover() override { delete Replacer; }
3279 /// \brief Really remove the instruction.
3280 void commit() override { delete Inst; }
3282 /// \brief Resurrect the instruction and reassign it to the proper uses if
3283 /// new value was provided when build this action.
3284 void undo() override {
3285 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n");
3286 Inserter.insert(Inst);
3294 /// Restoration point.
3295 /// The restoration point is a pointer to an action instead of an iterator
3296 /// because the iterator may be invalidated but not the pointer.
3297 typedef const TypePromotionAction *ConstRestorationPt;
3298 /// Advocate every changes made in that transaction.
3300 /// Undo all the changes made after the given point.
3301 void rollback(ConstRestorationPt Point);
3302 /// Get the current restoration point.
3303 ConstRestorationPt getRestorationPoint() const;
3305 /// \name API for IR modification with state keeping to support rollback.
3307 /// Same as Instruction::setOperand.
3308 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal);
3309 /// Same as Instruction::eraseFromParent.
3310 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr);
3311 /// Same as Value::replaceAllUsesWith.
3312 void replaceAllUsesWith(Instruction *Inst, Value *New);
3313 /// Same as Value::mutateType.
3314 void mutateType(Instruction *Inst, Type *NewTy);
3315 /// Same as IRBuilder::createTrunc.
3316 Value *createTrunc(Instruction *Opnd, Type *Ty);
3317 /// Same as IRBuilder::createSExt.
3318 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty);
3319 /// Same as IRBuilder::createZExt.
3320 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty);
3321 /// Same as Instruction::moveBefore.
3322 void moveBefore(Instruction *Inst, Instruction *Before);
3326 /// The ordered list of actions made so far.
3327 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions;
3328 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt;
3331 void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx,
3334 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal));
3337 void TypePromotionTransaction::eraseInstruction(Instruction *Inst,
3340 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal));
3343 void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst,
3345 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New));
3348 void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) {
3349 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy));
3352 Value *TypePromotionTransaction::createTrunc(Instruction *Opnd,
3354 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty));
3355 Value *Val = Ptr->getBuiltValue();
3356 Actions.push_back(std::move(Ptr));
3360 Value *TypePromotionTransaction::createSExt(Instruction *Inst,
3361 Value *Opnd, Type *Ty) {
3362 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty));
3363 Value *Val = Ptr->getBuiltValue();
3364 Actions.push_back(std::move(Ptr));
3368 Value *TypePromotionTransaction::createZExt(Instruction *Inst,
3369 Value *Opnd, Type *Ty) {
3370 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty));
3371 Value *Val = Ptr->getBuiltValue();
3372 Actions.push_back(std::move(Ptr));
3376 void TypePromotionTransaction::moveBefore(Instruction *Inst,
3377 Instruction *Before) {
3379 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before));
3382 TypePromotionTransaction::ConstRestorationPt
3383 TypePromotionTransaction::getRestorationPoint() const {
3384 return !Actions.empty() ? Actions.back().get() : nullptr;
3387 void TypePromotionTransaction::commit() {
3388 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt;
3394 void TypePromotionTransaction::rollback(
3395 TypePromotionTransaction::ConstRestorationPt Point) {
3396 while (!Actions.empty() && Point != Actions.back().get()) {
3397 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val();
3402 /// \brief A helper class for matching addressing modes.
3404 /// This encapsulates the logic for matching the target-legal addressing modes.
3405 class AddressingModeMatcher {
3406 SmallVectorImpl<Instruction*> &AddrModeInsts;
3407 const TargetMachine &TM;
3408 const TargetLowering &TLI;
3409 const DataLayout &DL;
3411 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
3412 /// the memory instruction that we're computing this address for.
3415 Instruction *MemoryInst;
3417 /// This is the addressing mode that we're building up. This is
3418 /// part of the return value of this addressing mode matching stuff.
3419 ExtAddrMode &AddrMode;
3421 /// The instructions inserted by other CodeGenPrepare optimizations.
3422 const SetOfInstrs &InsertedInsts;
3423 /// A map from the instructions to their type before promotion.
3424 InstrToOrigTy &PromotedInsts;
3425 /// The ongoing transaction where every action should be registered.
3426 TypePromotionTransaction &TPT;
3428 /// This is set to true when we should not do profitability checks.
3429 /// When true, IsProfitableToFoldIntoAddressingMode always returns true.
3430 bool IgnoreProfitability;
3432 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI,
3433 const TargetMachine &TM, Type *AT, unsigned AS,
3434 Instruction *MI, ExtAddrMode &AM,
3435 const SetOfInstrs &InsertedInsts,
3436 InstrToOrigTy &PromotedInsts,
3437 TypePromotionTransaction &TPT)
3438 : AddrModeInsts(AMI), TM(TM),
3439 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent())
3440 ->getTargetLowering()),
3441 DL(MI->getModule()->getDataLayout()), AccessTy(AT), AddrSpace(AS),
3442 MemoryInst(MI), AddrMode(AM), InsertedInsts(InsertedInsts),
3443 PromotedInsts(PromotedInsts), TPT(TPT) {
3444 IgnoreProfitability = false;
3448 /// Find the maximal addressing mode that a load/store of V can fold,
3449 /// give an access type of AccessTy. This returns a list of involved
3450 /// instructions in AddrModeInsts.
3451 /// \p InsertedInsts The instructions inserted by other CodeGenPrepare
3453 /// \p PromotedInsts maps the instructions to their type before promotion.
3454 /// \p The ongoing transaction where every action should be registered.
3455 static ExtAddrMode Match(Value *V, Type *AccessTy, unsigned AS,
3456 Instruction *MemoryInst,
3457 SmallVectorImpl<Instruction*> &AddrModeInsts,
3458 const TargetMachine &TM,
3459 const SetOfInstrs &InsertedInsts,
3460 InstrToOrigTy &PromotedInsts,
3461 TypePromotionTransaction &TPT) {
3464 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, AS,
3465 MemoryInst, Result, InsertedInsts,
3466 PromotedInsts, TPT).matchAddr(V, 0);
3467 (void)Success; assert(Success && "Couldn't select *anything*?");
3471 bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
3472 bool matchAddr(Value *V, unsigned Depth);
3473 bool matchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth,
3474 bool *MovedAway = nullptr);
3475 bool isProfitableToFoldIntoAddressingMode(Instruction *I,
3476 ExtAddrMode &AMBefore,
3477 ExtAddrMode &AMAfter);
3478 bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
3479 bool isPromotionProfitable(unsigned NewCost, unsigned OldCost,
3480 Value *PromotedOperand) const;
3483 /// Try adding ScaleReg*Scale to the current addressing mode.
3484 /// Return true and update AddrMode if this addr mode is legal for the target,
3486 bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale,
3488 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
3489 // mode. Just process that directly.
3491 return matchAddr(ScaleReg, Depth);
3493 // If the scale is 0, it takes nothing to add this.
3497 // If we already have a scale of this value, we can add to it, otherwise, we
3498 // need an available scale field.
3499 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
3502 ExtAddrMode TestAddrMode = AddrMode;
3504 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
3505 // [A+B + A*7] -> [B+A*8].
3506 TestAddrMode.Scale += Scale;
3507 TestAddrMode.ScaledReg = ScaleReg;
3509 // If the new address isn't legal, bail out.
3510 if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace))
3513 // It was legal, so commit it.
3514 AddrMode = TestAddrMode;
3516 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
3517 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
3518 // X*Scale + C*Scale to addr mode.
3519 ConstantInt *CI = nullptr; Value *AddLHS = nullptr;
3520 if (isa<Instruction>(ScaleReg) && // not a constant expr.
3521 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
3522 TestAddrMode.ScaledReg = AddLHS;
3523 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
3525 // If this addressing mode is legal, commit it and remember that we folded
3526 // this instruction.
3527 if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) {
3528 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
3529 AddrMode = TestAddrMode;
3534 // Otherwise, not (x+c)*scale, just return what we have.
3538 /// This is a little filter, which returns true if an addressing computation
3539 /// involving I might be folded into a load/store accessing it.
3540 /// This doesn't need to be perfect, but needs to accept at least
3541 /// the set of instructions that MatchOperationAddr can.
3542 static bool MightBeFoldableInst(Instruction *I) {
3543 switch (I->getOpcode()) {
3544 case Instruction::BitCast:
3545 case Instruction::AddrSpaceCast:
3546 // Don't touch identity bitcasts.
3547 if (I->getType() == I->getOperand(0)->getType())
3549 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
3550 case Instruction::PtrToInt:
3551 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3553 case Instruction::IntToPtr:
3554 // We know the input is intptr_t, so this is foldable.
3556 case Instruction::Add:
3558 case Instruction::Mul:
3559 case Instruction::Shl:
3560 // Can only handle X*C and X << C.
3561 return isa<ConstantInt>(I->getOperand(1));
3562 case Instruction::GetElementPtr:
3569 /// \brief Check whether or not \p Val is a legal instruction for \p TLI.
3570 /// \note \p Val is assumed to be the product of some type promotion.
3571 /// Therefore if \p Val has an undefined state in \p TLI, this is assumed
3572 /// to be legal, as the non-promoted value would have had the same state.
3573 static bool isPromotedInstructionLegal(const TargetLowering &TLI,
3574 const DataLayout &DL, Value *Val) {
3575 Instruction *PromotedInst = dyn_cast<Instruction>(Val);
3578 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode());
3579 // If the ISDOpcode is undefined, it was undefined before the promotion.
3582 // Otherwise, check if the promoted instruction is legal or not.
3583 return TLI.isOperationLegalOrCustom(
3584 ISDOpcode, TLI.getValueType(DL, PromotedInst->getType()));
3587 /// \brief Hepler class to perform type promotion.
3588 class TypePromotionHelper {
3589 /// \brief Utility function to check whether or not a sign or zero extension
3590 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by
3591 /// either using the operands of \p Inst or promoting \p Inst.
3592 /// The type of the extension is defined by \p IsSExt.
3593 /// In other words, check if:
3594 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType.
3595 /// #1 Promotion applies:
3596 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...).
3597 /// #2 Operand reuses:
3598 /// ext opnd1 to ConsideredExtType.
3599 /// \p PromotedInsts maps the instructions to their type before promotion.
3600 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType,
3601 const InstrToOrigTy &PromotedInsts, bool IsSExt);
3603 /// \brief Utility function to determine if \p OpIdx should be promoted when
3604 /// promoting \p Inst.
3605 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) {
3606 return !(isa<SelectInst>(Inst) && OpIdx == 0);
3609 /// \brief Utility function to promote the operand of \p Ext when this
3610 /// operand is a promotable trunc or sext or zext.
3611 /// \p PromotedInsts maps the instructions to their type before promotion.
3612 /// \p CreatedInstsCost[out] contains the cost of all instructions
3613 /// created to promote the operand of Ext.
3614 /// Newly added extensions are inserted in \p Exts.
3615 /// Newly added truncates are inserted in \p Truncs.
3616 /// Should never be called directly.
3617 /// \return The promoted value which is used instead of Ext.
3618 static Value *promoteOperandForTruncAndAnyExt(
3619 Instruction *Ext, TypePromotionTransaction &TPT,
3620 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3621 SmallVectorImpl<Instruction *> *Exts,
3622 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI);
3624 /// \brief Utility function to promote the operand of \p Ext when this
3625 /// operand is promotable and is not a supported trunc or sext.
3626 /// \p PromotedInsts maps the instructions to their type before promotion.
3627 /// \p CreatedInstsCost[out] contains the cost of all the instructions
3628 /// created to promote the operand of Ext.
3629 /// Newly added extensions are inserted in \p Exts.
3630 /// Newly added truncates are inserted in \p Truncs.
3631 /// Should never be called directly.
3632 /// \return The promoted value which is used instead of Ext.
3633 static Value *promoteOperandForOther(Instruction *Ext,
3634 TypePromotionTransaction &TPT,
3635 InstrToOrigTy &PromotedInsts,
3636 unsigned &CreatedInstsCost,
3637 SmallVectorImpl<Instruction *> *Exts,
3638 SmallVectorImpl<Instruction *> *Truncs,
3639 const TargetLowering &TLI, bool IsSExt);
3641 /// \see promoteOperandForOther.
3642 static Value *signExtendOperandForOther(
3643 Instruction *Ext, TypePromotionTransaction &TPT,
3644 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3645 SmallVectorImpl<Instruction *> *Exts,
3646 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3647 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3648 Exts, Truncs, TLI, true);
3651 /// \see promoteOperandForOther.
3652 static Value *zeroExtendOperandForOther(
3653 Instruction *Ext, TypePromotionTransaction &TPT,
3654 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3655 SmallVectorImpl<Instruction *> *Exts,
3656 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3657 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost,
3658 Exts, Truncs, TLI, false);
3662 /// Type for the utility function that promotes the operand of Ext.
3663 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT,
3664 InstrToOrigTy &PromotedInsts,
3665 unsigned &CreatedInstsCost,
3666 SmallVectorImpl<Instruction *> *Exts,
3667 SmallVectorImpl<Instruction *> *Truncs,
3668 const TargetLowering &TLI);
3669 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate
3670 /// action to promote the operand of \p Ext instead of using Ext.
3671 /// \return NULL if no promotable action is possible with the current
3673 /// \p InsertedInsts keeps track of all the instructions inserted by the
3674 /// other CodeGenPrepare optimizations. This information is important
3675 /// because we do not want to promote these instructions as CodeGenPrepare
3676 /// will reinsert them later. Thus creating an infinite loop: create/remove.
3677 /// \p PromotedInsts maps the instructions to their type before promotion.
3678 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts,
3679 const TargetLowering &TLI,
3680 const InstrToOrigTy &PromotedInsts);
3683 bool TypePromotionHelper::canGetThrough(const Instruction *Inst,
3684 Type *ConsideredExtType,
3685 const InstrToOrigTy &PromotedInsts,
3687 // The promotion helper does not know how to deal with vector types yet.
3688 // To be able to fix that, we would need to fix the places where we
3689 // statically extend, e.g., constants and such.
3690 if (Inst->getType()->isVectorTy())
3693 // We can always get through zext.
3694 if (isa<ZExtInst>(Inst))
3697 // sext(sext) is ok too.
3698 if (IsSExt && isa<SExtInst>(Inst))
3701 // We can get through binary operator, if it is legal. In other words, the
3702 // binary operator must have a nuw or nsw flag.
3703 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst);
3704 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) &&
3705 ((!IsSExt && BinOp->hasNoUnsignedWrap()) ||
3706 (IsSExt && BinOp->hasNoSignedWrap())))
3709 // Check if we can do the following simplification.
3710 // ext(trunc(opnd)) --> ext(opnd)
3711 if (!isa<TruncInst>(Inst))
3714 Value *OpndVal = Inst->getOperand(0);
3715 // Check if we can use this operand in the extension.
3716 // If the type is larger than the result type of the extension, we cannot.
3717 if (!OpndVal->getType()->isIntegerTy() ||
3718 OpndVal->getType()->getIntegerBitWidth() >
3719 ConsideredExtType->getIntegerBitWidth())
3722 // If the operand of the truncate is not an instruction, we will not have
3723 // any information on the dropped bits.
3724 // (Actually we could for constant but it is not worth the extra logic).
3725 Instruction *Opnd = dyn_cast<Instruction>(OpndVal);
3729 // Check if the source of the type is narrow enough.
3730 // I.e., check that trunc just drops extended bits of the same kind of
3732 // #1 get the type of the operand and check the kind of the extended bits.
3733 const Type *OpndType;
3734 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd);
3735 if (It != PromotedInsts.end() && It->second.getInt() == IsSExt)
3736 OpndType = It->second.getPointer();
3737 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd)))
3738 OpndType = Opnd->getOperand(0)->getType();
3742 // #2 check that the truncate just drops extended bits.
3743 return Inst->getType()->getIntegerBitWidth() >=
3744 OpndType->getIntegerBitWidth();
3747 TypePromotionHelper::Action TypePromotionHelper::getAction(
3748 Instruction *Ext, const SetOfInstrs &InsertedInsts,
3749 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) {
3750 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
3751 "Unexpected instruction type");
3752 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0));
3753 Type *ExtTy = Ext->getType();
3754 bool IsSExt = isa<SExtInst>(Ext);
3755 // If the operand of the extension is not an instruction, we cannot
3757 // If it, check we can get through.
3758 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt))
3761 // Do not promote if the operand has been added by codegenprepare.
3762 // Otherwise, it means we are undoing an optimization that is likely to be
3763 // redone, thus causing potential infinite loop.
3764 if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd))
3767 // SExt or Trunc instructions.
3768 // Return the related handler.
3769 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) ||
3770 isa<ZExtInst>(ExtOpnd))
3771 return promoteOperandForTruncAndAnyExt;
3773 // Regular instruction.
3774 // Abort early if we will have to insert non-free instructions.
3775 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType()))
3777 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther;
3780 Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt(
3781 llvm::Instruction *SExt, TypePromotionTransaction &TPT,
3782 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3783 SmallVectorImpl<Instruction *> *Exts,
3784 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) {
3785 // By construction, the operand of SExt is an instruction. Otherwise we cannot
3786 // get through it and this method should not be called.
3787 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0));
3788 Value *ExtVal = SExt;
3789 bool HasMergedNonFreeExt = false;
3790 if (isa<ZExtInst>(SExtOpnd)) {
3791 // Replace s|zext(zext(opnd))
3793 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd);
3795 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType());
3796 TPT.replaceAllUsesWith(SExt, ZExt);
3797 TPT.eraseInstruction(SExt);
3800 // Replace z|sext(trunc(opnd)) or sext(sext(opnd))
3802 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0));
3804 CreatedInstsCost = 0;
3806 // Remove dead code.
3807 if (SExtOpnd->use_empty())
3808 TPT.eraseInstruction(SExtOpnd);
3810 // Check if the extension is still needed.
3811 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal);
3812 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) {
3815 Exts->push_back(ExtInst);
3816 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt;
3821 // At this point we have: ext ty opnd to ty.
3822 // Reassign the uses of ExtInst to the opnd and remove ExtInst.
3823 Value *NextVal = ExtInst->getOperand(0);
3824 TPT.eraseInstruction(ExtInst, NextVal);
3828 Value *TypePromotionHelper::promoteOperandForOther(
3829 Instruction *Ext, TypePromotionTransaction &TPT,
3830 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost,
3831 SmallVectorImpl<Instruction *> *Exts,
3832 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI,
3834 // By construction, the operand of Ext is an instruction. Otherwise we cannot
3835 // get through it and this method should not be called.
3836 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0));
3837 CreatedInstsCost = 0;
3838 if (!ExtOpnd->hasOneUse()) {
3839 // ExtOpnd will be promoted.
3840 // All its uses, but Ext, will need to use a truncated value of the
3841 // promoted version.
3842 // Create the truncate now.
3843 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType());
3844 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) {
3845 ITrunc->removeFromParent();
3846 // Insert it just after the definition.
3847 ITrunc->insertAfter(ExtOpnd);
3849 Truncs->push_back(ITrunc);
3852 TPT.replaceAllUsesWith(ExtOpnd, Trunc);
3853 // Restore the operand of Ext (which has been replaced by the previous call
3854 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext.
3855 TPT.setOperand(Ext, 0, ExtOpnd);
3858 // Get through the Instruction:
3859 // 1. Update its type.
3860 // 2. Replace the uses of Ext by Inst.
3861 // 3. Extend each operand that needs to be extended.
3863 // Remember the original type of the instruction before promotion.
3864 // This is useful to know that the high bits are sign extended bits.
3865 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>(
3866 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt)));
3868 TPT.mutateType(ExtOpnd, Ext->getType());
3870 TPT.replaceAllUsesWith(Ext, ExtOpnd);
3872 Instruction *ExtForOpnd = Ext;
3874 DEBUG(dbgs() << "Propagate Ext to operands\n");
3875 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx;
3877 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n');
3878 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() ||
3879 !shouldExtOperand(ExtOpnd, OpIdx)) {
3880 DEBUG(dbgs() << "No need to propagate\n");
3883 // Check if we can statically extend the operand.
3884 Value *Opnd = ExtOpnd->getOperand(OpIdx);
3885 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) {
3886 DEBUG(dbgs() << "Statically extend\n");
3887 unsigned BitWidth = Ext->getType()->getIntegerBitWidth();
3888 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth)
3889 : Cst->getValue().zext(BitWidth);
3890 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal));
3893 // UndefValue are typed, so we have to statically sign extend them.
3894 if (isa<UndefValue>(Opnd)) {
3895 DEBUG(dbgs() << "Statically extend\n");
3896 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType()));
3900 // Otherwise we have to explicity sign extend the operand.
3901 // Check if Ext was reused to extend an operand.
3903 // If yes, create a new one.
3904 DEBUG(dbgs() << "More operands to ext\n");
3905 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType())
3906 : TPT.createZExt(Ext, Opnd, Ext->getType());
3907 if (!isa<Instruction>(ValForExtOpnd)) {
3908 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd);
3911 ExtForOpnd = cast<Instruction>(ValForExtOpnd);
3914 Exts->push_back(ExtForOpnd);
3915 TPT.setOperand(ExtForOpnd, 0, Opnd);
3917 // Move the sign extension before the insertion point.
3918 TPT.moveBefore(ExtForOpnd, ExtOpnd);
3919 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd);
3920 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd);
3921 // If more sext are required, new instructions will have to be created.
3922 ExtForOpnd = nullptr;
3924 if (ExtForOpnd == Ext) {
3925 DEBUG(dbgs() << "Extension is useless now\n");
3926 TPT.eraseInstruction(Ext);
3931 /// Check whether or not promoting an instruction to a wider type is profitable.
3932 /// \p NewCost gives the cost of extension instructions created by the
3934 /// \p OldCost gives the cost of extension instructions before the promotion
3935 /// plus the number of instructions that have been
3936 /// matched in the addressing mode the promotion.
3937 /// \p PromotedOperand is the value that has been promoted.
3938 /// \return True if the promotion is profitable, false otherwise.
3939 bool AddressingModeMatcher::isPromotionProfitable(
3940 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const {
3941 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n');
3942 // The cost of the new extensions is greater than the cost of the
3943 // old extension plus what we folded.
3944 // This is not profitable.
3945 if (NewCost > OldCost)
3947 if (NewCost < OldCost)
3949 // The promotion is neutral but it may help folding the sign extension in
3950 // loads for instance.
3951 // Check that we did not create an illegal instruction.
3952 return isPromotedInstructionLegal(TLI, DL, PromotedOperand);
3955 /// Given an instruction or constant expr, see if we can fold the operation
3956 /// into the addressing mode. If so, update the addressing mode and return
3957 /// true, otherwise return false without modifying AddrMode.
3958 /// If \p MovedAway is not NULL, it contains the information of whether or
3959 /// not AddrInst has to be folded into the addressing mode on success.
3960 /// If \p MovedAway == true, \p AddrInst will not be part of the addressing
3961 /// because it has been moved away.
3962 /// Thus AddrInst must not be added in the matched instructions.
3963 /// This state can happen when AddrInst is a sext, since it may be moved away.
3964 /// Therefore, AddrInst may not be valid when MovedAway is true and it must
3965 /// not be referenced anymore.
3966 bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode,
3969 // Avoid exponential behavior on extremely deep expression trees.
3970 if (Depth >= 5) return false;
3972 // By default, all matched instructions stay in place.
3977 case Instruction::PtrToInt:
3978 // PtrToInt is always a noop, as we know that the int type is pointer sized.
3979 return matchAddr(AddrInst->getOperand(0), Depth);
3980 case Instruction::IntToPtr: {
3981 auto AS = AddrInst->getType()->getPointerAddressSpace();
3982 auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS));
3983 // This inttoptr is a no-op if the integer type is pointer sized.
3984 if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy)
3985 return matchAddr(AddrInst->getOperand(0), Depth);
3988 case Instruction::BitCast:
3989 // BitCast is always a noop, and we can handle it as long as it is
3990 // int->int or pointer->pointer (we don't want int<->fp or something).
3991 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
3992 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
3993 // Don't touch identity bitcasts. These were probably put here by LSR,
3994 // and we don't want to mess around with them. Assume it knows what it
3996 AddrInst->getOperand(0)->getType() != AddrInst->getType())
3997 return matchAddr(AddrInst->getOperand(0), Depth);
3999 case Instruction::AddrSpaceCast: {
4001 = AddrInst->getOperand(0)->getType()->getPointerAddressSpace();
4002 unsigned DestAS = AddrInst->getType()->getPointerAddressSpace();
4003 if (TLI.isNoopAddrSpaceCast(SrcAS, DestAS))
4004 return matchAddr(AddrInst->getOperand(0), Depth);
4007 case Instruction::Add: {
4008 // Check to see if we can merge in the RHS then the LHS. If so, we win.
4009 ExtAddrMode BackupAddrMode = AddrMode;
4010 unsigned OldSize = AddrModeInsts.size();
4011 // Start a transaction at this point.
4012 // The LHS may match but not the RHS.
4013 // Therefore, we need a higher level restoration point to undo partially
4014 // matched operation.
4015 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4016 TPT.getRestorationPoint();
4018 if (matchAddr(AddrInst->getOperand(1), Depth+1) &&
4019 matchAddr(AddrInst->getOperand(0), Depth+1))
4022 // Restore the old addr mode info.
4023 AddrMode = BackupAddrMode;
4024 AddrModeInsts.resize(OldSize);
4025 TPT.rollback(LastKnownGood);
4027 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
4028 if (matchAddr(AddrInst->getOperand(0), Depth+1) &&
4029 matchAddr(AddrInst->getOperand(1), Depth+1))
4032 // Otherwise we definitely can't merge the ADD in.
4033 AddrMode = BackupAddrMode;
4034 AddrModeInsts.resize(OldSize);
4035 TPT.rollback(LastKnownGood);
4038 //case Instruction::Or:
4039 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
4041 case Instruction::Mul:
4042 case Instruction::Shl: {
4043 // Can only handle X*C and X << C.
4044 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
4047 int64_t Scale = RHS->getSExtValue();
4048 if (Opcode == Instruction::Shl)
4049 Scale = 1LL << Scale;
4051 return matchScaledValue(AddrInst->getOperand(0), Scale, Depth);
4053 case Instruction::GetElementPtr: {
4054 // Scan the GEP. We check it if it contains constant offsets and at most
4055 // one variable offset.
4056 int VariableOperand = -1;
4057 unsigned VariableScale = 0;
4059 int64_t ConstantOffset = 0;
4060 gep_type_iterator GTI = gep_type_begin(AddrInst);
4061 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
4062 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
4063 const StructLayout *SL = DL.getStructLayout(STy);
4065 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
4066 ConstantOffset += SL->getElementOffset(Idx);
4068 uint64_t TypeSize = DL.getTypeAllocSize(GTI.getIndexedType());
4069 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
4070 ConstantOffset += CI->getSExtValue()*TypeSize;
4071 } else if (TypeSize) { // Scales of zero don't do anything.
4072 // We only allow one variable index at the moment.
4073 if (VariableOperand != -1)
4076 // Remember the variable index.
4077 VariableOperand = i;
4078 VariableScale = TypeSize;
4083 // A common case is for the GEP to only do a constant offset. In this case,
4084 // just add it to the disp field and check validity.
4085 if (VariableOperand == -1) {
4086 AddrMode.BaseOffs += ConstantOffset;
4087 if (ConstantOffset == 0 ||
4088 TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) {
4089 // Check to see if we can fold the base pointer in too.
4090 if (matchAddr(AddrInst->getOperand(0), Depth+1))
4093 AddrMode.BaseOffs -= ConstantOffset;
4097 // Save the valid addressing mode in case we can't match.
4098 ExtAddrMode BackupAddrMode = AddrMode;
4099 unsigned OldSize = AddrModeInsts.size();
4101 // See if the scale and offset amount is valid for this target.
4102 AddrMode.BaseOffs += ConstantOffset;
4104 // Match the base operand of the GEP.
4105 if (!matchAddr(AddrInst->getOperand(0), Depth+1)) {
4106 // If it couldn't be matched, just stuff the value in a register.
4107 if (AddrMode.HasBaseReg) {
4108 AddrMode = BackupAddrMode;
4109 AddrModeInsts.resize(OldSize);
4112 AddrMode.HasBaseReg = true;
4113 AddrMode.BaseReg = AddrInst->getOperand(0);
4116 // Match the remaining variable portion of the GEP.
4117 if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
4119 // If it couldn't be matched, try stuffing the base into a register
4120 // instead of matching it, and retrying the match of the scale.
4121 AddrMode = BackupAddrMode;
4122 AddrModeInsts.resize(OldSize);
4123 if (AddrMode.HasBaseReg)
4125 AddrMode.HasBaseReg = true;
4126 AddrMode.BaseReg = AddrInst->getOperand(0);
4127 AddrMode.BaseOffs += ConstantOffset;
4128 if (!matchScaledValue(AddrInst->getOperand(VariableOperand),
4129 VariableScale, Depth)) {
4130 // If even that didn't work, bail.
4131 AddrMode = BackupAddrMode;
4132 AddrModeInsts.resize(OldSize);
4139 case Instruction::SExt:
4140 case Instruction::ZExt: {
4141 Instruction *Ext = dyn_cast<Instruction>(AddrInst);
4145 // Try to move this ext out of the way of the addressing mode.
4146 // Ask for a method for doing so.
4147 TypePromotionHelper::Action TPH =
4148 TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts);
4152 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4153 TPT.getRestorationPoint();
4154 unsigned CreatedInstsCost = 0;
4155 unsigned ExtCost = !TLI.isExtFree(Ext);
4156 Value *PromotedOperand =
4157 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI);
4158 // SExt has been moved away.
4159 // Thus either it will be rematched later in the recursive calls or it is
4160 // gone. Anyway, we must not fold it into the addressing mode at this point.
4164 // addr = gep base, idx
4166 // promotedOpnd = ext opnd <- no match here
4167 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls)
4168 // addr = gep base, op <- match
4172 assert(PromotedOperand &&
4173 "TypePromotionHelper should have filtered out those cases");
4175 ExtAddrMode BackupAddrMode = AddrMode;
4176 unsigned OldSize = AddrModeInsts.size();
4178 if (!matchAddr(PromotedOperand, Depth) ||
4179 // The total of the new cost is equal to the cost of the created
4181 // The total of the old cost is equal to the cost of the extension plus
4182 // what we have saved in the addressing mode.
4183 !isPromotionProfitable(CreatedInstsCost,
4184 ExtCost + (AddrModeInsts.size() - OldSize),
4186 AddrMode = BackupAddrMode;
4187 AddrModeInsts.resize(OldSize);
4188 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n");
4189 TPT.rollback(LastKnownGood);
4198 /// If we can, try to add the value of 'Addr' into the current addressing mode.
4199 /// If Addr can't be added to AddrMode this returns false and leaves AddrMode
4200 /// unmodified. This assumes that Addr is either a pointer type or intptr_t
4203 bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) {
4204 // Start a transaction at this point that we will rollback if the matching
4206 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4207 TPT.getRestorationPoint();
4208 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
4209 // Fold in immediates if legal for the target.
4210 AddrMode.BaseOffs += CI->getSExtValue();
4211 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4213 AddrMode.BaseOffs -= CI->getSExtValue();
4214 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
4215 // If this is a global variable, try to fold it into the addressing mode.
4216 if (!AddrMode.BaseGV) {
4217 AddrMode.BaseGV = GV;
4218 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4220 AddrMode.BaseGV = nullptr;
4222 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
4223 ExtAddrMode BackupAddrMode = AddrMode;
4224 unsigned OldSize = AddrModeInsts.size();
4226 // Check to see if it is possible to fold this operation.
4227 bool MovedAway = false;
4228 if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) {
4229 // This instruction may have been moved away. If so, there is nothing
4233 // Okay, it's possible to fold this. Check to see if it is actually
4234 // *profitable* to do so. We use a simple cost model to avoid increasing
4235 // register pressure too much.
4236 if (I->hasOneUse() ||
4237 isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
4238 AddrModeInsts.push_back(I);
4242 // It isn't profitable to do this, roll back.
4243 //cerr << "NOT FOLDING: " << *I;
4244 AddrMode = BackupAddrMode;
4245 AddrModeInsts.resize(OldSize);
4246 TPT.rollback(LastKnownGood);
4248 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
4249 if (matchOperationAddr(CE, CE->getOpcode(), Depth))
4251 TPT.rollback(LastKnownGood);
4252 } else if (isa<ConstantPointerNull>(Addr)) {
4253 // Null pointer gets folded without affecting the addressing mode.
4257 // Worse case, the target should support [reg] addressing modes. :)
4258 if (!AddrMode.HasBaseReg) {
4259 AddrMode.HasBaseReg = true;
4260 AddrMode.BaseReg = Addr;
4261 // Still check for legality in case the target supports [imm] but not [i+r].
4262 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4264 AddrMode.HasBaseReg = false;
4265 AddrMode.BaseReg = nullptr;
4268 // If the base register is already taken, see if we can do [r+r].
4269 if (AddrMode.Scale == 0) {
4271 AddrMode.ScaledReg = Addr;
4272 if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace))
4275 AddrMode.ScaledReg = nullptr;
4278 TPT.rollback(LastKnownGood);
4282 /// Check to see if all uses of OpVal by the specified inline asm call are due
4283 /// to memory operands. If so, return true, otherwise return false.
4284 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
4285 const TargetMachine &TM) {
4286 const Function *F = CI->getParent()->getParent();
4287 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering();
4288 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo();
4289 TargetLowering::AsmOperandInfoVector TargetConstraints =
4290 TLI->ParseConstraints(F->getParent()->getDataLayout(), TRI,
4291 ImmutableCallSite(CI));
4292 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4293 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4295 // Compute the constraint code and ConstraintType to use.
4296 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4298 // If this asm operand is our Value*, and if it isn't an indirect memory
4299 // operand, we can't fold it!
4300 if (OpInfo.CallOperandVal == OpVal &&
4301 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
4302 !OpInfo.isIndirect))
4309 /// Recursively walk all the uses of I until we find a memory use.
4310 /// If we find an obviously non-foldable instruction, return true.
4311 /// Add the ultimately found memory instructions to MemoryUses.
4312 static bool FindAllMemoryUses(
4314 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses,
4315 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) {
4316 // If we already considered this instruction, we're done.
4317 if (!ConsideredInsts.insert(I).second)
4320 // If this is an obviously unfoldable instruction, bail out.
4321 if (!MightBeFoldableInst(I))
4324 // Loop over all the uses, recursively processing them.
4325 for (Use &U : I->uses()) {
4326 Instruction *UserI = cast<Instruction>(U.getUser());
4328 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) {
4329 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo()));
4333 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) {
4334 unsigned opNo = U.getOperandNo();
4335 if (opNo == 0) return true; // Storing addr, not into addr.
4336 MemoryUses.push_back(std::make_pair(SI, opNo));
4340 if (CallInst *CI = dyn_cast<CallInst>(UserI)) {
4341 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
4342 if (!IA) return true;
4344 // If this is a memory operand, we're cool, otherwise bail out.
4345 if (!IsOperandAMemoryOperand(CI, IA, I, TM))
4350 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM))
4357 /// Return true if Val is already known to be live at the use site that we're
4358 /// folding it into. If so, there is no cost to include it in the addressing
4359 /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the
4360 /// instruction already.
4361 bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
4362 Value *KnownLive2) {
4363 // If Val is either of the known-live values, we know it is live!
4364 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2)
4367 // All values other than instructions and arguments (e.g. constants) are live.
4368 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
4370 // If Val is a constant sized alloca in the entry block, it is live, this is
4371 // true because it is just a reference to the stack/frame pointer, which is
4372 // live for the whole function.
4373 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
4374 if (AI->isStaticAlloca())
4377 // Check to see if this value is already used in the memory instruction's
4378 // block. If so, it's already live into the block at the very least, so we
4379 // can reasonably fold it.
4380 return Val->isUsedInBasicBlock(MemoryInst->getParent());
4383 /// It is possible for the addressing mode of the machine to fold the specified
4384 /// instruction into a load or store that ultimately uses it.
4385 /// However, the specified instruction has multiple uses.
4386 /// Given this, it may actually increase register pressure to fold it
4387 /// into the load. For example, consider this code:
4391 /// use(Y) -> nonload/store
4395 /// In this case, Y has multiple uses, and can be folded into the load of Z
4396 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
4397 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
4398 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
4399 /// number of computations either.
4401 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
4402 /// X was live across 'load Z' for other reasons, we actually *would* want to
4403 /// fold the addressing mode in the Z case. This would make Y die earlier.
4404 bool AddressingModeMatcher::
4405 isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
4406 ExtAddrMode &AMAfter) {
4407 if (IgnoreProfitability) return true;
4409 // AMBefore is the addressing mode before this instruction was folded into it,
4410 // and AMAfter is the addressing mode after the instruction was folded. Get
4411 // the set of registers referenced by AMAfter and subtract out those
4412 // referenced by AMBefore: this is the set of values which folding in this
4413 // address extends the lifetime of.
4415 // Note that there are only two potential values being referenced here,
4416 // BaseReg and ScaleReg (global addresses are always available, as are any
4417 // folded immediates).
4418 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
4420 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
4421 // lifetime wasn't extended by adding this instruction.
4422 if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4424 if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
4425 ScaledReg = nullptr;
4427 // If folding this instruction (and it's subexprs) didn't extend any live
4428 // ranges, we're ok with it.
4429 if (!BaseReg && !ScaledReg)
4432 // If all uses of this instruction are ultimately load/store/inlineasm's,
4433 // check to see if their addressing modes will include this instruction. If
4434 // so, we can fold it into all uses, so it doesn't matter if it has multiple
4436 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
4437 SmallPtrSet<Instruction*, 16> ConsideredInsts;
4438 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM))
4439 return false; // Has a non-memory, non-foldable use!
4441 // Now that we know that all uses of this instruction are part of a chain of
4442 // computation involving only operations that could theoretically be folded
4443 // into a memory use, loop over each of these uses and see if they could
4444 // *actually* fold the instruction.
4445 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
4446 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
4447 Instruction *User = MemoryUses[i].first;
4448 unsigned OpNo = MemoryUses[i].second;
4450 // Get the access type of this use. If the use isn't a pointer, we don't
4451 // know what it accesses.
4452 Value *Address = User->getOperand(OpNo);
4453 PointerType *AddrTy = dyn_cast<PointerType>(Address->getType());
4456 Type *AddressAccessTy = AddrTy->getElementType();
4457 unsigned AS = AddrTy->getAddressSpace();
4459 // Do a match against the root of this address, ignoring profitability. This
4460 // will tell us if the addressing mode for the memory operation will
4461 // *actually* cover the shared instruction.
4463 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4464 TPT.getRestorationPoint();
4465 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, AS,
4466 MemoryInst, Result, InsertedInsts,
4467 PromotedInsts, TPT);
4468 Matcher.IgnoreProfitability = true;
4469 bool Success = Matcher.matchAddr(Address, 0);
4470 (void)Success; assert(Success && "Couldn't select *anything*?");
4472 // The match was to check the profitability, the changes made are not
4473 // part of the original matcher. Therefore, they should be dropped
4474 // otherwise the original matcher will not present the right state.
4475 TPT.rollback(LastKnownGood);
4477 // If the match didn't cover I, then it won't be shared by it.
4478 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
4479 I) == MatchedAddrModeInsts.end())
4482 MatchedAddrModeInsts.clear();
4488 } // end anonymous namespace
4490 /// Return true if the specified values are defined in a
4491 /// different basic block than BB.
4492 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
4493 if (Instruction *I = dyn_cast<Instruction>(V))
4494 return I->getParent() != BB;
4498 /// Load and Store Instructions often have addressing modes that can do
4499 /// significant amounts of computation. As such, instruction selection will try
4500 /// to get the load or store to do as much computation as possible for the
4501 /// program. The problem is that isel can only see within a single block. As
4502 /// such, we sink as much legal addressing mode work into the block as possible.
4504 /// This method is used to optimize both load/store and inline asms with memory
4506 bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
4507 Type *AccessTy, unsigned AddrSpace) {
4510 // Try to collapse single-value PHI nodes. This is necessary to undo
4511 // unprofitable PRE transformations.
4512 SmallVector<Value*, 8> worklist;
4513 SmallPtrSet<Value*, 16> Visited;
4514 worklist.push_back(Addr);
4516 // Use a worklist to iteratively look through PHI nodes, and ensure that
4517 // the addressing mode obtained from the non-PHI roots of the graph
4519 Value *Consensus = nullptr;
4520 unsigned NumUsesConsensus = 0;
4521 bool IsNumUsesConsensusValid = false;
4522 SmallVector<Instruction*, 16> AddrModeInsts;
4523 ExtAddrMode AddrMode;
4524 TypePromotionTransaction TPT;
4525 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4526 TPT.getRestorationPoint();
4527 while (!worklist.empty()) {
4528 Value *V = worklist.back();
4529 worklist.pop_back();
4531 // Break use-def graph loops.
4532 if (!Visited.insert(V).second) {
4533 Consensus = nullptr;
4537 // For a PHI node, push all of its incoming values.
4538 if (PHINode *P = dyn_cast<PHINode>(V)) {
4539 for (Value *IncValue : P->incoming_values())
4540 worklist.push_back(IncValue);
4544 // For non-PHIs, determine the addressing mode being computed.
4545 SmallVector<Instruction*, 16> NewAddrModeInsts;
4546 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match(
4547 V, AccessTy, AddrSpace, MemoryInst, NewAddrModeInsts, *TM,
4548 InsertedInsts, PromotedInsts, TPT);
4550 // This check is broken into two cases with very similar code to avoid using
4551 // getNumUses() as much as possible. Some values have a lot of uses, so
4552 // calling getNumUses() unconditionally caused a significant compile-time
4556 AddrMode = NewAddrMode;
4557 AddrModeInsts = NewAddrModeInsts;
4559 } else if (NewAddrMode == AddrMode) {
4560 if (!IsNumUsesConsensusValid) {
4561 NumUsesConsensus = Consensus->getNumUses();
4562 IsNumUsesConsensusValid = true;
4565 // Ensure that the obtained addressing mode is equivalent to that obtained
4566 // for all other roots of the PHI traversal. Also, when choosing one
4567 // such root as representative, select the one with the most uses in order
4568 // to keep the cost modeling heuristics in AddressingModeMatcher
4570 unsigned NumUses = V->getNumUses();
4571 if (NumUses > NumUsesConsensus) {
4573 NumUsesConsensus = NumUses;
4574 AddrModeInsts = NewAddrModeInsts;
4579 Consensus = nullptr;
4583 // If the addressing mode couldn't be determined, or if multiple different
4584 // ones were determined, bail out now.
4586 TPT.rollback(LastKnownGood);
4591 // Check to see if any of the instructions supersumed by this addr mode are
4592 // non-local to I's BB.
4593 bool AnyNonLocal = false;
4594 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
4595 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
4601 // If all the instructions matched are already in this BB, don't do anything.
4603 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
4607 // Insert this computation right after this user. Since our caller is
4608 // scanning from the top of the BB to the bottom, reuse of the expr are
4609 // guaranteed to happen later.
4610 IRBuilder<> Builder(MemoryInst);
4612 // Now that we determined the addressing expression we want to use and know
4613 // that we have to sink it into this block. Check to see if we have already
4614 // done this for some other load/store instr in this block. If so, reuse the
4616 Value *&SunkAddr = SunkAddrs[Addr];
4618 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
4619 << *MemoryInst << "\n");
4620 if (SunkAddr->getType() != Addr->getType())
4621 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4622 } else if (AddrSinkUsingGEPs ||
4623 (!AddrSinkUsingGEPs.getNumOccurrences() && TM &&
4624 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent())
4626 // By default, we use the GEP-based method when AA is used later. This
4627 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities.
4628 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4629 << *MemoryInst << "\n");
4630 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4631 Value *ResultPtr = nullptr, *ResultIndex = nullptr;
4633 // First, find the pointer.
4634 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) {
4635 ResultPtr = AddrMode.BaseReg;
4636 AddrMode.BaseReg = nullptr;
4639 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) {
4640 // We can't add more than one pointer together, nor can we scale a
4641 // pointer (both of which seem meaningless).
4642 if (ResultPtr || AddrMode.Scale != 1)
4645 ResultPtr = AddrMode.ScaledReg;
4649 if (AddrMode.BaseGV) {
4653 ResultPtr = AddrMode.BaseGV;
4656 // If the real base value actually came from an inttoptr, then the matcher
4657 // will look through it and provide only the integer value. In that case,
4659 if (!ResultPtr && AddrMode.BaseReg) {
4661 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr");
4662 AddrMode.BaseReg = nullptr;
4663 } else if (!ResultPtr && AddrMode.Scale == 1) {
4665 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr");
4670 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) {
4671 SunkAddr = Constant::getNullValue(Addr->getType());
4672 } else if (!ResultPtr) {
4676 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace());
4677 Type *I8Ty = Builder.getInt8Ty();
4679 // Start with the base register. Do this first so that subsequent address
4680 // matching finds it last, which will prevent it from trying to match it
4681 // as the scaled value in case it happens to be a mul. That would be
4682 // problematic if we've sunk a different mul for the scale, because then
4683 // we'd end up sinking both muls.
4684 if (AddrMode.BaseReg) {
4685 Value *V = AddrMode.BaseReg;
4686 if (V->getType() != IntPtrTy)
4687 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4692 // Add the scale value.
4693 if (AddrMode.Scale) {
4694 Value *V = AddrMode.ScaledReg;
4695 if (V->getType() == IntPtrTy) {
4697 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4698 cast<IntegerType>(V->getType())->getBitWidth()) {
4699 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4701 // It is only safe to sign extend the BaseReg if we know that the math
4702 // required to create it did not overflow before we extend it. Since
4703 // the original IR value was tossed in favor of a constant back when
4704 // the AddrMode was created we need to bail out gracefully if widths
4705 // do not match instead of extending it.
4706 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex);
4707 if (I && (ResultIndex != AddrMode.BaseReg))
4708 I->eraseFromParent();
4712 if (AddrMode.Scale != 1)
4713 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4716 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr");
4721 // Add in the Base Offset if present.
4722 if (AddrMode.BaseOffs) {
4723 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4725 // We need to add this separately from the scale above to help with
4726 // SDAG consecutive load/store merging.
4727 if (ResultPtr->getType() != I8PtrTy)
4728 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4729 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4736 SunkAddr = ResultPtr;
4738 if (ResultPtr->getType() != I8PtrTy)
4739 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy);
4740 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr");
4743 if (SunkAddr->getType() != Addr->getType())
4744 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
4747 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
4748 << *MemoryInst << "\n");
4749 Type *IntPtrTy = DL->getIntPtrType(Addr->getType());
4750 Value *Result = nullptr;
4752 // Start with the base register. Do this first so that subsequent address
4753 // matching finds it last, which will prevent it from trying to match it
4754 // as the scaled value in case it happens to be a mul. That would be
4755 // problematic if we've sunk a different mul for the scale, because then
4756 // we'd end up sinking both muls.
4757 if (AddrMode.BaseReg) {
4758 Value *V = AddrMode.BaseReg;
4759 if (V->getType()->isPointerTy())
4760 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4761 if (V->getType() != IntPtrTy)
4762 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
4766 // Add the scale value.
4767 if (AddrMode.Scale) {
4768 Value *V = AddrMode.ScaledReg;
4769 if (V->getType() == IntPtrTy) {
4771 } else if (V->getType()->isPointerTy()) {
4772 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
4773 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
4774 cast<IntegerType>(V->getType())->getBitWidth()) {
4775 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
4777 // It is only safe to sign extend the BaseReg if we know that the math
4778 // required to create it did not overflow before we extend it. Since
4779 // the original IR value was tossed in favor of a constant back when
4780 // the AddrMode was created we need to bail out gracefully if widths
4781 // do not match instead of extending it.
4782 Instruction *I = dyn_cast_or_null<Instruction>(Result);
4783 if (I && (Result != AddrMode.BaseReg))
4784 I->eraseFromParent();
4787 if (AddrMode.Scale != 1)
4788 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
4791 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4796 // Add in the BaseGV if present.
4797 if (AddrMode.BaseGV) {
4798 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
4800 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4805 // Add in the Base Offset if present.
4806 if (AddrMode.BaseOffs) {
4807 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
4809 Result = Builder.CreateAdd(Result, V, "sunkaddr");
4815 SunkAddr = Constant::getNullValue(Addr->getType());
4817 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
4820 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
4822 // If we have no uses, recursively delete the value and all dead instructions
4824 if (Repl->use_empty()) {
4825 // This can cause recursive deletion, which can invalidate our iterator.
4826 // Use a WeakVH to hold onto it in case this happens.
4827 WeakVH IterHandle(&*CurInstIterator);
4828 BasicBlock *BB = CurInstIterator->getParent();
4830 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
4832 if (IterHandle != CurInstIterator.getNodePtrUnchecked()) {
4833 // If the iterator instruction was recursively deleted, start over at the
4834 // start of the block.
4835 CurInstIterator = BB->begin();
4843 /// If there are any memory operands, use OptimizeMemoryInst to sink their
4844 /// address computing into the block when possible / profitable.
4845 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
4846 bool MadeChange = false;
4848 const TargetRegisterInfo *TRI =
4849 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo();
4850 TargetLowering::AsmOperandInfoVector TargetConstraints =
4851 TLI->ParseConstraints(*DL, TRI, CS);
4853 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
4854 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
4856 // Compute the constraint code and ConstraintType to use.
4857 TLI->ComputeConstraintToUse(OpInfo, SDValue());
4859 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4860 OpInfo.isIndirect) {
4861 Value *OpVal = CS->getArgOperand(ArgNo++);
4862 MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u);
4863 } else if (OpInfo.Type == InlineAsm::isInput)
4870 /// \brief Check if all the uses of \p Inst are equivalent (or free) zero or
4871 /// sign extensions.
4872 static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) {
4873 assert(!Inst->use_empty() && "Input must have at least one use");
4874 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin());
4875 bool IsSExt = isa<SExtInst>(FirstUser);
4876 Type *ExtTy = FirstUser->getType();
4877 for (const User *U : Inst->users()) {
4878 const Instruction *UI = cast<Instruction>(U);
4879 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI)))
4881 Type *CurTy = UI->getType();
4882 // Same input and output types: Same instruction after CSE.
4886 // If IsSExt is true, we are in this situation:
4888 // b = sext ty1 a to ty2
4889 // c = sext ty1 a to ty3
4890 // Assuming ty2 is shorter than ty3, this could be turned into:
4892 // b = sext ty1 a to ty2
4893 // c = sext ty2 b to ty3
4894 // However, the last sext is not free.
4898 // This is a ZExt, maybe this is free to extend from one type to another.
4899 // In that case, we would not account for a different use.
4902 if (ExtTy->getScalarType()->getIntegerBitWidth() >
4903 CurTy->getScalarType()->getIntegerBitWidth()) {
4911 if (!TLI.isZExtFree(NarrowTy, LargeTy))
4914 // All uses are the same or can be derived from one another for free.
4918 /// \brief Try to form ExtLd by promoting \p Exts until they reach a
4919 /// load instruction.
4920 /// If an ext(load) can be formed, it is returned via \p LI for the load
4921 /// and \p Inst for the extension.
4922 /// Otherwise LI == nullptr and Inst == nullptr.
4923 /// When some promotion happened, \p TPT contains the proper state to
4926 /// \return true when promoting was necessary to expose the ext(load)
4927 /// opportunity, false otherwise.
4931 /// %ld = load i32* %addr
4932 /// %add = add nuw i32 %ld, 4
4933 /// %zext = zext i32 %add to i64
4937 /// %ld = load i32* %addr
4938 /// %zext = zext i32 %ld to i64
4939 /// %add = add nuw i64 %zext, 4
4941 /// Thanks to the promotion, we can match zext(load i32*) to i64.
4942 bool CodeGenPrepare::extLdPromotion(TypePromotionTransaction &TPT,
4943 LoadInst *&LI, Instruction *&Inst,
4944 const SmallVectorImpl<Instruction *> &Exts,
4945 unsigned CreatedInstsCost = 0) {
4946 // Iterate over all the extensions to see if one form an ext(load).
4947 for (auto I : Exts) {
4948 // Check if we directly have ext(load).
4949 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) {
4951 // No promotion happened here.
4954 // Check whether or not we want to do any promotion.
4955 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion)
4957 // Get the action to perform the promotion.
4958 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
4959 I, InsertedInsts, *TLI, PromotedInsts);
4960 // Check if we can promote.
4963 // Save the current state.
4964 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
4965 TPT.getRestorationPoint();
4966 SmallVector<Instruction *, 4> NewExts;
4967 unsigned NewCreatedInstsCost = 0;
4968 unsigned ExtCost = !TLI->isExtFree(I);
4970 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost,
4971 &NewExts, nullptr, *TLI);
4972 assert(PromotedVal &&
4973 "TypePromotionHelper should have filtered out those cases");
4975 // We would be able to merge only one extension in a load.
4976 // Therefore, if we have more than 1 new extension we heuristically
4977 // cut this search path, because it means we degrade the code quality.
4978 // With exactly 2, the transformation is neutral, because we will merge
4979 // one extension but leave one. However, we optimistically keep going,
4980 // because the new extension may be removed too.
4981 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost;
4982 TotalCreatedInstsCost -= ExtCost;
4983 if (!StressExtLdPromotion &&
4984 (TotalCreatedInstsCost > 1 ||
4985 !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) {
4986 // The promotion is not profitable, rollback to the previous state.
4987 TPT.rollback(LastKnownGood);
4990 // The promotion is profitable.
4991 // Check if it exposes an ext(load).
4992 (void)extLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost);
4993 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost ||
4994 // If we have created a new extension, i.e., now we have two
4995 // extensions. We must make sure one of them is merged with
4996 // the load, otherwise we may degrade the code quality.
4997 (LI->hasOneUse() || hasSameExtUse(LI, *TLI))))
4998 // Promotion happened.
5000 // If this does not help to expose an ext(load) then, rollback.
5001 TPT.rollback(LastKnownGood);
5003 // None of the extension can form an ext(load).
5009 /// Move a zext or sext fed by a load into the same basic block as the load,
5010 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
5011 /// extend into the load.
5012 /// \p I[in/out] the extension may be modified during the process if some
5013 /// promotions apply.
5015 bool CodeGenPrepare::moveExtToFormExtLoad(Instruction *&I) {
5016 // Try to promote a chain of computation if it allows to form
5017 // an extended load.
5018 TypePromotionTransaction TPT;
5019 TypePromotionTransaction::ConstRestorationPt LastKnownGood =
5020 TPT.getRestorationPoint();
5021 SmallVector<Instruction *, 1> Exts;
5023 // Look for a load being extended.
5024 LoadInst *LI = nullptr;
5025 Instruction *OldExt = I;
5026 bool HasPromoted = extLdPromotion(TPT, LI, I, Exts);
5028 assert(!HasPromoted && !LI && "If we did not match any load instruction "
5029 "the code must remain the same");
5034 // If they're already in the same block, there's nothing to do.
5035 // Make the cheap checks first if we did not promote.
5036 // If we promoted, we need to check if it is indeed profitable.
5037 if (!HasPromoted && LI->getParent() == I->getParent())
5040 EVT VT = TLI->getValueType(*DL, I->getType());
5041 EVT LoadVT = TLI->getValueType(*DL, LI->getType());
5043 // If the load has other users and the truncate is not free, this probably
5044 // isn't worthwhile.
5045 if (!LI->hasOneUse() && TLI &&
5046 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
5047 !TLI->isTruncateFree(I->getType(), LI->getType())) {
5049 TPT.rollback(LastKnownGood);
5053 // Check whether the target supports casts folded into loads.
5055 if (isa<ZExtInst>(I))
5056 LType = ISD::ZEXTLOAD;
5058 assert(isa<SExtInst>(I) && "Unexpected ext type!");
5059 LType = ISD::SEXTLOAD;
5061 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) {
5063 TPT.rollback(LastKnownGood);
5067 // Move the extend into the same block as the load, so that SelectionDAG
5070 I->removeFromParent();
5076 bool CodeGenPrepare::optimizeExtUses(Instruction *I) {
5077 BasicBlock *DefBB = I->getParent();
5079 // If the result of a {s|z}ext and its source are both live out, rewrite all
5080 // other uses of the source with result of extension.
5081 Value *Src = I->getOperand(0);
5082 if (Src->hasOneUse())
5085 // Only do this xform if truncating is free.
5086 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
5089 // Only safe to perform the optimization if the source is also defined in
5091 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
5094 bool DefIsLiveOut = false;
5095 for (User *U : I->users()) {
5096 Instruction *UI = cast<Instruction>(U);
5098 // Figure out which BB this ext is used in.
5099 BasicBlock *UserBB = UI->getParent();
5100 if (UserBB == DefBB) continue;
5101 DefIsLiveOut = true;
5107 // Make sure none of the uses are PHI nodes.
5108 for (User *U : Src->users()) {
5109 Instruction *UI = cast<Instruction>(U);
5110 BasicBlock *UserBB = UI->getParent();
5111 if (UserBB == DefBB) continue;
5112 // Be conservative. We don't want this xform to end up introducing
5113 // reloads just before load / store instructions.
5114 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI))
5118 // InsertedTruncs - Only insert one trunc in each block once.
5119 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
5121 bool MadeChange = false;
5122 for (Use &U : Src->uses()) {
5123 Instruction *User = cast<Instruction>(U.getUser());
5125 // Figure out which BB this ext is used in.
5126 BasicBlock *UserBB = User->getParent();
5127 if (UserBB == DefBB) continue;
5129 // Both src and def are live in this block. Rewrite the use.
5130 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
5132 if (!InsertedTrunc) {
5133 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5134 assert(InsertPt != UserBB->end());
5135 InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt);
5136 InsertedInsts.insert(InsertedTrunc);
5139 // Replace a use of the {s|z}ext source with a use of the result.
5148 // Find loads whose uses only use some of the loaded value's bits. Add an "and"
5149 // just after the load if the target can fold this into one extload instruction,
5150 // with the hope of eliminating some of the other later "and" instructions using
5151 // the loaded value. "and"s that are made trivially redundant by the insertion
5152 // of the new "and" are removed by this function, while others (e.g. those whose
5153 // path from the load goes through a phi) are left for isel to potentially
5186 // becomes (after a call to optimizeLoadExt for each load):
5190 // x1' = and x1, 0xff
5194 // x2' = and x2, 0xff
5201 bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) {
5203 if (!Load->isSimple() ||
5204 !(Load->getType()->isIntegerTy() || Load->getType()->isPointerTy()))
5207 // Skip loads we've already transformed or have no reason to transform.
5208 if (Load->hasOneUse()) {
5209 User *LoadUser = *Load->user_begin();
5210 if (cast<Instruction>(LoadUser)->getParent() == Load->getParent() &&
5211 !dyn_cast<PHINode>(LoadUser))
5215 // Look at all uses of Load, looking through phis, to determine how many bits
5216 // of the loaded value are needed.
5217 SmallVector<Instruction *, 8> WorkList;
5218 SmallPtrSet<Instruction *, 16> Visited;
5219 SmallVector<Instruction *, 8> AndsToMaybeRemove;
5220 for (auto *U : Load->users())
5221 WorkList.push_back(cast<Instruction>(U));
5223 EVT LoadResultVT = TLI->getValueType(*DL, Load->getType());
5224 unsigned BitWidth = LoadResultVT.getSizeInBits();
5225 APInt DemandBits(BitWidth, 0);
5226 APInt WidestAndBits(BitWidth, 0);
5228 while (!WorkList.empty()) {
5229 Instruction *I = WorkList.back();
5230 WorkList.pop_back();
5232 // Break use-def graph loops.
5233 if (!Visited.insert(I).second)
5236 // For a PHI node, push all of its users.
5237 if (auto *Phi = dyn_cast<PHINode>(I)) {
5238 for (auto *U : Phi->users())
5239 WorkList.push_back(cast<Instruction>(U));
5243 switch (I->getOpcode()) {
5244 case llvm::Instruction::And: {
5245 auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1));
5248 APInt AndBits = AndC->getValue();
5249 DemandBits |= AndBits;
5250 // Keep track of the widest and mask we see.
5251 if (AndBits.ugt(WidestAndBits))
5252 WidestAndBits = AndBits;
5253 if (AndBits == WidestAndBits && I->getOperand(0) == Load)
5254 AndsToMaybeRemove.push_back(I);
5258 case llvm::Instruction::Shl: {
5259 auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1));
5262 uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1);
5263 auto ShlDemandBits = APInt::getAllOnesValue(BitWidth).lshr(ShiftAmt);
5264 DemandBits |= ShlDemandBits;
5268 case llvm::Instruction::Trunc: {
5269 EVT TruncVT = TLI->getValueType(*DL, I->getType());
5270 unsigned TruncBitWidth = TruncVT.getSizeInBits();
5271 auto TruncBits = APInt::getAllOnesValue(TruncBitWidth).zext(BitWidth);
5272 DemandBits |= TruncBits;
5281 uint32_t ActiveBits = DemandBits.getActiveBits();
5282 // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the
5283 // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example,
5284 // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but
5285 // (and (load x) 1) is not matched as a single instruction, rather as a LDR
5286 // followed by an AND.
5287 // TODO: Look into removing this restriction by fixing backends to either
5288 // return false for isLoadExtLegal for i1 or have them select this pattern to
5289 // a single instruction.
5291 // Also avoid hoisting if we didn't see any ands with the exact DemandBits
5292 // mask, since these are the only ands that will be removed by isel.
5293 if (ActiveBits <= 1 || !APIntOps::isMask(ActiveBits, DemandBits) ||
5294 WidestAndBits != DemandBits)
5297 LLVMContext &Ctx = Load->getType()->getContext();
5298 Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits);
5299 EVT TruncVT = TLI->getValueType(*DL, TruncTy);
5301 // Reject cases that won't be matched as extloads.
5302 if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() ||
5303 !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT))
5306 IRBuilder<> Builder(Load->getNextNode());
5307 auto *NewAnd = dyn_cast<Instruction>(
5308 Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits)));
5310 // Replace all uses of load with new and (except for the use of load in the
5312 Load->replaceAllUsesWith(NewAnd);
5313 NewAnd->setOperand(0, Load);
5315 // Remove any and instructions that are now redundant.
5316 for (auto *And : AndsToMaybeRemove)
5317 // Check that the and mask is the same as the one we decided to put on the
5319 if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) {
5320 And->replaceAllUsesWith(NewAnd);
5321 if (&*CurInstIterator == And)
5322 CurInstIterator = std::next(And->getIterator());
5323 And->eraseFromParent();
5331 /// Check if V (an operand of a select instruction) is an expensive instruction
5332 /// that is only used once.
5333 static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) {
5334 auto *I = dyn_cast<Instruction>(V);
5335 // If it's safe to speculatively execute, then it should not have side
5336 // effects; therefore, it's safe to sink and possibly *not* execute.
5337 return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) &&
5338 TTI->getUserCost(I) >= TargetTransformInfo::TCC_Expensive;
5341 /// Returns true if a SelectInst should be turned into an explicit branch.
5342 static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI,
5344 // FIXME: This should use the same heuristics as IfConversion to determine
5345 // whether a select is better represented as a branch. This requires that
5346 // branch probability metadata is preserved for the select, which is not the
5349 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
5351 // If a branch is predictable, an out-of-order CPU can avoid blocking on its
5352 // comparison condition. If the compare has more than one use, there's
5353 // probably another cmov or setcc around, so it's not worth emitting a branch.
5354 if (!Cmp || !Cmp->hasOneUse())
5357 Value *CmpOp0 = Cmp->getOperand(0);
5358 Value *CmpOp1 = Cmp->getOperand(1);
5360 // Emit "cmov on compare with a memory operand" as a branch to avoid stalls
5361 // on a load from memory. But if the load is used more than once, do not
5362 // change the select to a branch because the load is probably needed
5363 // regardless of whether the branch is taken or not.
5364 if ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
5365 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()))
5368 // If either operand of the select is expensive and only needed on one side
5369 // of the select, we should form a branch.
5370 if (sinkSelectOperand(TTI, SI->getTrueValue()) ||
5371 sinkSelectOperand(TTI, SI->getFalseValue()))
5378 /// If we have a SelectInst that will likely profit from branch prediction,
5379 /// turn it into a branch.
5380 bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
5381 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
5383 // Can we convert the 'select' to CF ?
5384 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
5387 TargetLowering::SelectSupportKind SelectKind;
5389 SelectKind = TargetLowering::VectorMaskSelect;
5390 else if (SI->getType()->isVectorTy())
5391 SelectKind = TargetLowering::ScalarCondVectorVal;
5393 SelectKind = TargetLowering::ScalarValSelect;
5395 // Do we have efficient codegen support for this kind of 'selects' ?
5396 if (TLI->isSelectSupported(SelectKind)) {
5397 // We have efficient codegen support for the select instruction.
5398 // Check if it is profitable to keep this 'select'.
5399 if (!TLI->isPredictableSelectExpensive() ||
5400 !isFormingBranchFromSelectProfitable(TTI, SI))
5406 // Transform a sequence like this:
5408 // %cmp = cmp uge i32 %a, %b
5409 // %sel = select i1 %cmp, i32 %c, i32 %d
5413 // %cmp = cmp uge i32 %a, %b
5414 // br i1 %cmp, label %select.true, label %select.false
5416 // br label %select.end
5418 // br label %select.end
5420 // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ]
5422 // In addition, we may sink instructions that produce %c or %d from
5423 // the entry block into the destination(s) of the new branch.
5424 // If the true or false blocks do not contain a sunken instruction, that
5425 // block and its branch may be optimized away. In that case, one side of the
5426 // first branch will point directly to select.end, and the corresponding PHI
5427 // predecessor block will be the start block.
5429 // First, we split the block containing the select into 2 blocks.
5430 BasicBlock *StartBlock = SI->getParent();
5431 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
5432 BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
5434 // Delete the unconditional branch that was just created by the split.
5435 StartBlock->getTerminator()->eraseFromParent();
5437 // These are the new basic blocks for the conditional branch.
5438 // At least one will become an actual new basic block.
5439 BasicBlock *TrueBlock = nullptr;
5440 BasicBlock *FalseBlock = nullptr;
5442 // Sink expensive instructions into the conditional blocks to avoid executing
5443 // them speculatively.
5444 if (sinkSelectOperand(TTI, SI->getTrueValue())) {
5445 TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink",
5446 EndBlock->getParent(), EndBlock);
5447 auto *TrueBranch = BranchInst::Create(EndBlock, TrueBlock);
5448 auto *TrueInst = cast<Instruction>(SI->getTrueValue());
5449 TrueInst->moveBefore(TrueBranch);
5451 if (sinkSelectOperand(TTI, SI->getFalseValue())) {
5452 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink",
5453 EndBlock->getParent(), EndBlock);
5454 auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock);
5455 auto *FalseInst = cast<Instruction>(SI->getFalseValue());
5456 FalseInst->moveBefore(FalseBranch);
5459 // If there was nothing to sink, then arbitrarily choose the 'false' side
5460 // for a new input value to the PHI.
5461 if (TrueBlock == FalseBlock) {
5462 assert(TrueBlock == nullptr &&
5463 "Unexpected basic block transform while optimizing select");
5465 FalseBlock = BasicBlock::Create(SI->getContext(), "select.false",
5466 EndBlock->getParent(), EndBlock);
5467 BranchInst::Create(EndBlock, FalseBlock);
5470 // Insert the real conditional branch based on the original condition.
5471 // If we did not create a new block for one of the 'true' or 'false' paths
5472 // of the condition, it means that side of the branch goes to the end block
5473 // directly and the path originates from the start block from the point of
5474 // view of the new PHI.
5475 if (TrueBlock == nullptr) {
5476 BranchInst::Create(EndBlock, FalseBlock, SI->getCondition(), SI);
5477 TrueBlock = StartBlock;
5478 } else if (FalseBlock == nullptr) {
5479 BranchInst::Create(TrueBlock, EndBlock, SI->getCondition(), SI);
5480 FalseBlock = StartBlock;
5482 BranchInst::Create(TrueBlock, FalseBlock, SI->getCondition(), SI);
5485 // The select itself is replaced with a PHI Node.
5486 PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front());
5488 PN->addIncoming(SI->getTrueValue(), TrueBlock);
5489 PN->addIncoming(SI->getFalseValue(), FalseBlock);
5491 SI->replaceAllUsesWith(PN);
5492 SI->eraseFromParent();
5494 // Instruct OptimizeBlock to skip to the next block.
5495 CurInstIterator = StartBlock->end();
5496 ++NumSelectsExpanded;
5500 static bool isBroadcastShuffle(ShuffleVectorInst *SVI) {
5501 SmallVector<int, 16> Mask(SVI->getShuffleMask());
5503 for (unsigned i = 0; i < Mask.size(); ++i) {
5504 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem)
5506 SplatElem = Mask[i];
5512 /// Some targets have expensive vector shifts if the lanes aren't all the same
5513 /// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases
5514 /// it's often worth sinking a shufflevector splat down to its use so that
5515 /// codegen can spot all lanes are identical.
5516 bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) {
5517 BasicBlock *DefBB = SVI->getParent();
5519 // Only do this xform if variable vector shifts are particularly expensive.
5520 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType()))
5523 // We only expect better codegen by sinking a shuffle if we can recognise a
5525 if (!isBroadcastShuffle(SVI))
5528 // InsertedShuffles - Only insert a shuffle in each block once.
5529 DenseMap<BasicBlock*, Instruction*> InsertedShuffles;
5531 bool MadeChange = false;
5532 for (User *U : SVI->users()) {
5533 Instruction *UI = cast<Instruction>(U);
5535 // Figure out which BB this ext is used in.
5536 BasicBlock *UserBB = UI->getParent();
5537 if (UserBB == DefBB) continue;
5539 // For now only apply this when the splat is used by a shift instruction.
5540 if (!UI->isShift()) continue;
5542 // Everything checks out, sink the shuffle if the user's block doesn't
5543 // already have a copy.
5544 Instruction *&InsertedShuffle = InsertedShuffles[UserBB];
5546 if (!InsertedShuffle) {
5547 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
5548 assert(InsertPt != UserBB->end());
5550 new ShuffleVectorInst(SVI->getOperand(0), SVI->getOperand(1),
5551 SVI->getOperand(2), "", &*InsertPt);
5554 UI->replaceUsesOfWith(SVI, InsertedShuffle);
5558 // If we removed all uses, nuke the shuffle.
5559 if (SVI->use_empty()) {
5560 SVI->eraseFromParent();
5567 bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) {
5571 Value *Cond = SI->getCondition();
5572 Type *OldType = Cond->getType();
5573 LLVMContext &Context = Cond->getContext();
5574 MVT RegType = TLI->getRegisterType(Context, TLI->getValueType(*DL, OldType));
5575 unsigned RegWidth = RegType.getSizeInBits();
5577 if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth())
5580 // If the register width is greater than the type width, expand the condition
5581 // of the switch instruction and each case constant to the width of the
5582 // register. By widening the type of the switch condition, subsequent
5583 // comparisons (for case comparisons) will not need to be extended to the
5584 // preferred register width, so we will potentially eliminate N-1 extends,
5585 // where N is the number of cases in the switch.
5586 auto *NewType = Type::getIntNTy(Context, RegWidth);
5588 // Zero-extend the switch condition and case constants unless the switch
5589 // condition is a function argument that is already being sign-extended.
5590 // In that case, we can avoid an unnecessary mask/extension by sign-extending
5591 // everything instead.
5592 Instruction::CastOps ExtType = Instruction::ZExt;
5593 if (auto *Arg = dyn_cast<Argument>(Cond))
5594 if (Arg->hasSExtAttr())
5595 ExtType = Instruction::SExt;
5597 auto *ExtInst = CastInst::Create(ExtType, Cond, NewType);
5598 ExtInst->insertBefore(SI);
5599 SI->setCondition(ExtInst);
5600 for (SwitchInst::CaseIt Case : SI->cases()) {
5601 APInt NarrowConst = Case.getCaseValue()->getValue();
5602 APInt WideConst = (ExtType == Instruction::ZExt) ?
5603 NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth);
5604 Case.setValue(ConstantInt::get(Context, WideConst));
5611 /// \brief Helper class to promote a scalar operation to a vector one.
5612 /// This class is used to move downward extractelement transition.
5614 /// a = vector_op <2 x i32>
5615 /// b = extractelement <2 x i32> a, i32 0
5620 /// a = vector_op <2 x i32>
5621 /// c = vector_op a (equivalent to scalar_op on the related lane)
5622 /// * d = extractelement <2 x i32> c, i32 0
5624 /// Assuming both extractelement and store can be combine, we get rid of the
5626 class VectorPromoteHelper {
5627 /// DataLayout associated with the current module.
5628 const DataLayout &DL;
5630 /// Used to perform some checks on the legality of vector operations.
5631 const TargetLowering &TLI;
5633 /// Used to estimated the cost of the promoted chain.
5634 const TargetTransformInfo &TTI;
5636 /// The transition being moved downwards.
5637 Instruction *Transition;
5638 /// The sequence of instructions to be promoted.
5639 SmallVector<Instruction *, 4> InstsToBePromoted;
5640 /// Cost of combining a store and an extract.
5641 unsigned StoreExtractCombineCost;
5642 /// Instruction that will be combined with the transition.
5643 Instruction *CombineInst;
5645 /// \brief The instruction that represents the current end of the transition.
5646 /// Since we are faking the promotion until we reach the end of the chain
5647 /// of computation, we need a way to get the current end of the transition.
5648 Instruction *getEndOfTransition() const {
5649 if (InstsToBePromoted.empty())
5651 return InstsToBePromoted.back();
5654 /// \brief Return the index of the original value in the transition.
5655 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value,
5656 /// c, is at index 0.
5657 unsigned getTransitionOriginalValueIdx() const {
5658 assert(isa<ExtractElementInst>(Transition) &&
5659 "Other kind of transitions are not supported yet");
5663 /// \brief Return the index of the index in the transition.
5664 /// E.g., for "extractelement <2 x i32> c, i32 0" the index
5666 unsigned getTransitionIdx() const {
5667 assert(isa<ExtractElementInst>(Transition) &&
5668 "Other kind of transitions are not supported yet");
5672 /// \brief Get the type of the transition.
5673 /// This is the type of the original value.
5674 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the
5675 /// transition is <2 x i32>.
5676 Type *getTransitionType() const {
5677 return Transition->getOperand(getTransitionOriginalValueIdx())->getType();
5680 /// \brief Promote \p ToBePromoted by moving \p Def downward through.
5681 /// I.e., we have the following sequence:
5682 /// Def = Transition <ty1> a to <ty2>
5683 /// b = ToBePromoted <ty2> Def, ...
5685 /// b = ToBePromoted <ty1> a, ...
5686 /// Def = Transition <ty1> ToBePromoted to <ty2>
5687 void promoteImpl(Instruction *ToBePromoted);
5689 /// \brief Check whether or not it is profitable to promote all the
5690 /// instructions enqueued to be promoted.
5691 bool isProfitableToPromote() {
5692 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx());
5693 unsigned Index = isa<ConstantInt>(ValIdx)
5694 ? cast<ConstantInt>(ValIdx)->getZExtValue()
5696 Type *PromotedType = getTransitionType();
5698 StoreInst *ST = cast<StoreInst>(CombineInst);
5699 unsigned AS = ST->getPointerAddressSpace();
5700 unsigned Align = ST->getAlignment();
5701 // Check if this store is supported.
5702 if (!TLI.allowsMisalignedMemoryAccesses(
5703 TLI.getValueType(DL, ST->getValueOperand()->getType()), AS,
5705 // If this is not supported, there is no way we can combine
5706 // the extract with the store.
5710 // The scalar chain of computation has to pay for the transition
5711 // scalar to vector.
5712 // The vector chain has to account for the combining cost.
5713 uint64_t ScalarCost =
5714 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index);
5715 uint64_t VectorCost = StoreExtractCombineCost;
5716 for (const auto &Inst : InstsToBePromoted) {
5717 // Compute the cost.
5718 // By construction, all instructions being promoted are arithmetic ones.
5719 // Moreover, one argument is a constant that can be viewed as a splat
5721 Value *Arg0 = Inst->getOperand(0);
5722 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) ||
5723 isa<ConstantFP>(Arg0);
5724 TargetTransformInfo::OperandValueKind Arg0OVK =
5725 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5726 : TargetTransformInfo::OK_AnyValue;
5727 TargetTransformInfo::OperandValueKind Arg1OVK =
5728 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue
5729 : TargetTransformInfo::OK_AnyValue;
5730 ScalarCost += TTI.getArithmeticInstrCost(
5731 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK);
5732 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType,
5735 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: "
5736 << ScalarCost << "\nVector: " << VectorCost << '\n');
5737 return ScalarCost > VectorCost;
5740 /// \brief Generate a constant vector with \p Val with the same
5741 /// number of elements as the transition.
5742 /// \p UseSplat defines whether or not \p Val should be replicated
5743 /// across the whole vector.
5744 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>,
5745 /// otherwise we generate a vector with as many undef as possible:
5746 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only
5747 /// used at the index of the extract.
5748 Value *getConstantVector(Constant *Val, bool UseSplat) const {
5749 unsigned ExtractIdx = UINT_MAX;
5751 // If we cannot determine where the constant must be, we have to
5752 // use a splat constant.
5753 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx());
5754 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx))
5755 ExtractIdx = CstVal->getSExtValue();
5760 unsigned End = getTransitionType()->getVectorNumElements();
5762 return ConstantVector::getSplat(End, Val);
5764 SmallVector<Constant *, 4> ConstVec;
5765 UndefValue *UndefVal = UndefValue::get(Val->getType());
5766 for (unsigned Idx = 0; Idx != End; ++Idx) {
5767 if (Idx == ExtractIdx)
5768 ConstVec.push_back(Val);
5770 ConstVec.push_back(UndefVal);
5772 return ConstantVector::get(ConstVec);
5775 /// \brief Check if promoting to a vector type an operand at \p OperandIdx
5776 /// in \p Use can trigger undefined behavior.
5777 static bool canCauseUndefinedBehavior(const Instruction *Use,
5778 unsigned OperandIdx) {
5779 // This is not safe to introduce undef when the operand is on
5780 // the right hand side of a division-like instruction.
5781 if (OperandIdx != 1)
5783 switch (Use->getOpcode()) {
5786 case Instruction::SDiv:
5787 case Instruction::UDiv:
5788 case Instruction::SRem:
5789 case Instruction::URem:
5791 case Instruction::FDiv:
5792 case Instruction::FRem:
5793 return !Use->hasNoNaNs();
5795 llvm_unreachable(nullptr);
5799 VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI,
5800 const TargetTransformInfo &TTI, Instruction *Transition,
5801 unsigned CombineCost)
5802 : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition),
5803 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) {
5804 assert(Transition && "Do not know how to promote null");
5807 /// \brief Check if we can promote \p ToBePromoted to \p Type.
5808 bool canPromote(const Instruction *ToBePromoted) const {
5809 // We could support CastInst too.
5810 return isa<BinaryOperator>(ToBePromoted);
5813 /// \brief Check if it is profitable to promote \p ToBePromoted
5814 /// by moving downward the transition through.
5815 bool shouldPromote(const Instruction *ToBePromoted) const {
5816 // Promote only if all the operands can be statically expanded.
5817 // Indeed, we do not want to introduce any new kind of transitions.
5818 for (const Use &U : ToBePromoted->operands()) {
5819 const Value *Val = U.get();
5820 if (Val == getEndOfTransition()) {
5821 // If the use is a division and the transition is on the rhs,
5822 // we cannot promote the operation, otherwise we may create a
5823 // division by zero.
5824 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()))
5828 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) &&
5829 !isa<ConstantFP>(Val))
5832 // Check that the resulting operation is legal.
5833 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode());
5836 return StressStoreExtract ||
5837 TLI.isOperationLegalOrCustom(
5838 ISDOpcode, TLI.getValueType(DL, getTransitionType(), true));
5841 /// \brief Check whether or not \p Use can be combined
5842 /// with the transition.
5843 /// I.e., is it possible to do Use(Transition) => AnotherUse?
5844 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); }
5846 /// \brief Record \p ToBePromoted as part of the chain to be promoted.
5847 void enqueueForPromotion(Instruction *ToBePromoted) {
5848 InstsToBePromoted.push_back(ToBePromoted);
5851 /// \brief Set the instruction that will be combined with the transition.
5852 void recordCombineInstruction(Instruction *ToBeCombined) {
5853 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine");
5854 CombineInst = ToBeCombined;
5857 /// \brief Promote all the instructions enqueued for promotion if it is
5859 /// \return True if the promotion happened, false otherwise.
5861 // Check if there is something to promote.
5862 // Right now, if we do not have anything to combine with,
5863 // we assume the promotion is not profitable.
5864 if (InstsToBePromoted.empty() || !CombineInst)
5868 if (!StressStoreExtract && !isProfitableToPromote())
5872 for (auto &ToBePromoted : InstsToBePromoted)
5873 promoteImpl(ToBePromoted);
5874 InstsToBePromoted.clear();
5878 } // End of anonymous namespace.
5880 void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) {
5881 // At this point, we know that all the operands of ToBePromoted but Def
5882 // can be statically promoted.
5883 // For Def, we need to use its parameter in ToBePromoted:
5884 // b = ToBePromoted ty1 a
5885 // Def = Transition ty1 b to ty2
5886 // Move the transition down.
5887 // 1. Replace all uses of the promoted operation by the transition.
5888 // = ... b => = ... Def.
5889 assert(ToBePromoted->getType() == Transition->getType() &&
5890 "The type of the result of the transition does not match "
5892 ToBePromoted->replaceAllUsesWith(Transition);
5893 // 2. Update the type of the uses.
5894 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def.
5895 Type *TransitionTy = getTransitionType();
5896 ToBePromoted->mutateType(TransitionTy);
5897 // 3. Update all the operands of the promoted operation with promoted
5899 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a.
5900 for (Use &U : ToBePromoted->operands()) {
5901 Value *Val = U.get();
5902 Value *NewVal = nullptr;
5903 if (Val == Transition)
5904 NewVal = Transition->getOperand(getTransitionOriginalValueIdx());
5905 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) ||
5906 isa<ConstantFP>(Val)) {
5907 // Use a splat constant if it is not safe to use undef.
5908 NewVal = getConstantVector(
5909 cast<Constant>(Val),
5910 isa<UndefValue>(Val) ||
5911 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo()));
5913 llvm_unreachable("Did you modified shouldPromote and forgot to update "
5915 ToBePromoted->setOperand(U.getOperandNo(), NewVal);
5917 Transition->removeFromParent();
5918 Transition->insertAfter(ToBePromoted);
5919 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted);
5922 /// Some targets can do store(extractelement) with one instruction.
5923 /// Try to push the extractelement towards the stores when the target
5924 /// has this feature and this is profitable.
5925 bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) {
5926 unsigned CombineCost = UINT_MAX;
5927 if (DisableStoreExtract || !TLI ||
5928 (!StressStoreExtract &&
5929 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(),
5930 Inst->getOperand(1), CombineCost)))
5933 // At this point we know that Inst is a vector to scalar transition.
5934 // Try to move it down the def-use chain, until:
5935 // - We can combine the transition with its single use
5936 // => we got rid of the transition.
5937 // - We escape the current basic block
5938 // => we would need to check that we are moving it at a cheaper place and
5939 // we do not do that for now.
5940 BasicBlock *Parent = Inst->getParent();
5941 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n');
5942 VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost);
5943 // If the transition has more than one use, assume this is not going to be
5945 while (Inst->hasOneUse()) {
5946 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin());
5947 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n');
5949 if (ToBePromoted->getParent() != Parent) {
5950 DEBUG(dbgs() << "Instruction to promote is in a different block ("
5951 << ToBePromoted->getParent()->getName()
5952 << ") than the transition (" << Parent->getName() << ").\n");
5956 if (VPH.canCombine(ToBePromoted)) {
5957 DEBUG(dbgs() << "Assume " << *Inst << '\n'
5958 << "will be combined with: " << *ToBePromoted << '\n');
5959 VPH.recordCombineInstruction(ToBePromoted);
5960 bool Changed = VPH.promote();
5961 NumStoreExtractExposed += Changed;
5965 DEBUG(dbgs() << "Try promoting.\n");
5966 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted))
5969 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n");
5971 VPH.enqueueForPromotion(ToBePromoted);
5972 Inst = ToBePromoted;
5977 bool CodeGenPrepare::optimizeInst(Instruction *I, bool& ModifiedDT) {
5978 // Bail out if we inserted the instruction to prevent optimizations from
5979 // stepping on each other's toes.
5980 if (InsertedInsts.count(I))
5983 if (PHINode *P = dyn_cast<PHINode>(I)) {
5984 // It is possible for very late stage optimizations (such as SimplifyCFG)
5985 // to introduce PHI nodes too late to be cleaned up. If we detect such a
5986 // trivial PHI, go ahead and zap it here.
5987 if (Value *V = SimplifyInstruction(P, *DL, TLInfo, nullptr)) {
5988 P->replaceAllUsesWith(V);
5989 P->eraseFromParent();
5996 if (CastInst *CI = dyn_cast<CastInst>(I)) {
5997 // If the source of the cast is a constant, then this should have
5998 // already been constant folded. The only reason NOT to constant fold
5999 // it is if something (e.g. LSR) was careful to place the constant
6000 // evaluation in a block other than then one that uses it (e.g. to hoist
6001 // the address of globals out of a loop). If this is the case, we don't
6002 // want to forward-subst the cast.
6003 if (isa<Constant>(CI->getOperand(0)))
6006 if (TLI && OptimizeNoopCopyExpression(CI, *TLI, *DL))
6009 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6010 /// Sink a zext or sext into its user blocks if the target type doesn't
6011 /// fit in one register
6013 TLI->getTypeAction(CI->getContext(),
6014 TLI->getValueType(*DL, CI->getType())) ==
6015 TargetLowering::TypeExpandInteger) {
6016 return SinkCast(CI);
6018 bool MadeChange = moveExtToFormExtLoad(I);
6019 return MadeChange | optimizeExtUses(I);
6025 if (CmpInst *CI = dyn_cast<CmpInst>(I))
6026 if (!TLI || !TLI->hasMultipleConditionRegisters())
6027 return OptimizeCmpExpression(CI);
6029 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
6030 stripInvariantGroupMetadata(*LI);
6032 bool Modified = optimizeLoadExt(LI);
6033 unsigned AS = LI->getPointerAddressSpace();
6034 Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS);
6040 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
6041 stripInvariantGroupMetadata(*SI);
6043 unsigned AS = SI->getPointerAddressSpace();
6044 return optimizeMemoryInst(I, SI->getOperand(1),
6045 SI->getOperand(0)->getType(), AS);
6050 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I);
6052 if (BinOp && (BinOp->getOpcode() == Instruction::AShr ||
6053 BinOp->getOpcode() == Instruction::LShr)) {
6054 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1));
6055 if (TLI && CI && TLI->hasExtractBitsInsn())
6056 return OptimizeExtractBits(BinOp, CI, *TLI, *DL);
6061 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
6062 if (GEPI->hasAllZeroIndices()) {
6063 /// The GEP operand must be a pointer, so must its result -> BitCast
6064 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
6065 GEPI->getName(), GEPI);
6066 GEPI->replaceAllUsesWith(NC);
6067 GEPI->eraseFromParent();
6069 optimizeInst(NC, ModifiedDT);
6075 if (CallInst *CI = dyn_cast<CallInst>(I))
6076 return optimizeCallInst(CI, ModifiedDT);
6078 if (SelectInst *SI = dyn_cast<SelectInst>(I))
6079 return optimizeSelectInst(SI);
6081 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I))
6082 return optimizeShuffleVectorInst(SVI);
6084 if (auto *Switch = dyn_cast<SwitchInst>(I))
6085 return optimizeSwitchInst(Switch);
6087 if (isa<ExtractElementInst>(I))
6088 return optimizeExtractElementInst(I);
6093 /// Given an OR instruction, check to see if this is a bitreverse
6094 /// idiom. If so, insert the new intrinsic and return true.
6095 static bool makeBitReverse(Instruction &I, const DataLayout &DL,
6096 const TargetLowering &TLI) {
6097 if (!I.getType()->isIntegerTy() ||
6098 !TLI.isOperationLegalOrCustom(ISD::BITREVERSE,
6099 TLI.getValueType(DL, I.getType(), true)))
6102 SmallVector<Instruction*, 4> Insts;
6103 if (!recognizeBitReverseOrBSwapIdiom(&I, false, true, Insts))
6105 Instruction *LastInst = Insts.back();
6106 I.replaceAllUsesWith(LastInst);
6107 RecursivelyDeleteTriviallyDeadInstructions(&I);
6111 // In this pass we look for GEP and cast instructions that are used
6112 // across basic blocks and rewrite them to improve basic-block-at-a-time
6114 bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool& ModifiedDT) {
6116 bool MadeChange = false;
6118 CurInstIterator = BB.begin();
6119 while (CurInstIterator != BB.end()) {
6120 MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT);
6125 bool MadeBitReverse = true;
6126 while (TLI && MadeBitReverse) {
6127 MadeBitReverse = false;
6128 for (auto &I : reverse(BB)) {
6129 if (makeBitReverse(I, *DL, *TLI)) {
6130 MadeBitReverse = MadeChange = true;
6135 MadeChange |= dupRetToEnableTailCallOpts(&BB);
6140 // llvm.dbg.value is far away from the value then iSel may not be able
6141 // handle it properly. iSel will drop llvm.dbg.value if it can not
6142 // find a node corresponding to the value.
6143 bool CodeGenPrepare::placeDbgValues(Function &F) {
6144 bool MadeChange = false;
6145 for (BasicBlock &BB : F) {
6146 Instruction *PrevNonDbgInst = nullptr;
6147 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
6148 Instruction *Insn = &*BI++;
6149 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
6150 // Leave dbg.values that refer to an alloca alone. These
6151 // instrinsics describe the address of a variable (= the alloca)
6152 // being taken. They should not be moved next to the alloca
6153 // (and to the beginning of the scope), but rather stay close to
6154 // where said address is used.
6155 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) {
6156 PrevNonDbgInst = Insn;
6160 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
6161 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
6162 // If VI is a phi in a block with an EHPad terminator, we can't insert
6164 if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad())
6166 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
6167 DVI->removeFromParent();
6168 if (isa<PHINode>(VI))
6169 DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt());
6171 DVI->insertAfter(VI);
6180 // If there is a sequence that branches based on comparing a single bit
6181 // against zero that can be combined into a single instruction, and the
6182 // target supports folding these into a single instruction, sink the
6183 // mask and compare into the branch uses. Do this before OptimizeBlock ->
6184 // OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being
6186 bool CodeGenPrepare::sinkAndCmp(Function &F) {
6187 if (!EnableAndCmpSinking)
6189 if (!TLI || !TLI->isMaskAndBranchFoldingLegal())
6191 bool MadeChange = false;
6192 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
6193 BasicBlock *BB = &*I++;
6195 // Does this BB end with the following?
6196 // %andVal = and %val, #single-bit-set
6197 // %icmpVal = icmp %andResult, 0
6198 // br i1 %cmpVal label %dest1, label %dest2"
6199 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator());
6200 if (!Brcc || !Brcc->isConditional())
6202 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0));
6203 if (!Cmp || Cmp->getParent() != BB)
6205 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1));
6206 if (!Zero || !Zero->isZero())
6208 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0));
6209 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB)
6211 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1));
6212 if (!Mask || !Mask->getUniqueInteger().isPowerOf2())
6214 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump());
6216 // Push the "and; icmp" for any users that are conditional branches.
6217 // Since there can only be one branch use per BB, we don't need to keep
6218 // track of which BBs we insert into.
6219 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end();
6223 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI);
6225 if (!BrccUser || !BrccUser->isConditional())
6227 BasicBlock *UserBB = BrccUser->getParent();
6228 if (UserBB == BB) continue;
6229 DEBUG(dbgs() << "found Brcc use\n");
6231 // Sink the "and; icmp" to use.
6233 BinaryOperator *NewAnd =
6234 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "",
6237 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero,
6241 DEBUG(BrccUser->getParent()->dump());
6247 /// \brief Retrieve the probabilities of a conditional branch. Returns true on
6248 /// success, or returns false if no or invalid metadata was found.
6249 static bool extractBranchMetadata(BranchInst *BI,
6250 uint64_t &ProbTrue, uint64_t &ProbFalse) {
6251 assert(BI->isConditional() &&
6252 "Looking for probabilities on unconditional branch?");
6253 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
6254 if (!ProfileData || ProfileData->getNumOperands() != 3)
6257 const auto *CITrue =
6258 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1));
6259 const auto *CIFalse =
6260 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2));
6261 if (!CITrue || !CIFalse)
6264 ProbTrue = CITrue->getValue().getZExtValue();
6265 ProbFalse = CIFalse->getValue().getZExtValue();
6270 /// \brief Scale down both weights to fit into uint32_t.
6271 static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) {
6272 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse;
6273 uint32_t Scale = (NewMax / UINT32_MAX) + 1;
6274 NewTrue = NewTrue / Scale;
6275 NewFalse = NewFalse / Scale;
6278 /// \brief Some targets prefer to split a conditional branch like:
6280 /// %0 = icmp ne i32 %a, 0
6281 /// %1 = icmp ne i32 %b, 0
6282 /// %or.cond = or i1 %0, %1
6283 /// br i1 %or.cond, label %TrueBB, label %FalseBB
6285 /// into multiple branch instructions like:
6288 /// %0 = icmp ne i32 %a, 0
6289 /// br i1 %0, label %TrueBB, label %bb2
6291 /// %1 = icmp ne i32 %b, 0
6292 /// br i1 %1, label %TrueBB, label %FalseBB
6294 /// This usually allows instruction selection to do even further optimizations
6295 /// and combine the compare with the branch instruction. Currently this is
6296 /// applied for targets which have "cheap" jump instructions.
6298 /// FIXME: Remove the (equivalent?) implementation in SelectionDAG.
6300 bool CodeGenPrepare::splitBranchCondition(Function &F) {
6301 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive())
6304 bool MadeChange = false;
6305 for (auto &BB : F) {
6306 // Does this BB end with the following?
6307 // %cond1 = icmp|fcmp|binary instruction ...
6308 // %cond2 = icmp|fcmp|binary instruction ...
6309 // %cond.or = or|and i1 %cond1, cond2
6310 // br i1 %cond.or label %dest1, label %dest2"
6311 BinaryOperator *LogicOp;
6312 BasicBlock *TBB, *FBB;
6313 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB)))
6316 auto *Br1 = cast<BranchInst>(BB.getTerminator());
6317 if (Br1->getMetadata(LLVMContext::MD_unpredictable))
6321 Value *Cond1, *Cond2;
6322 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)),
6323 m_OneUse(m_Value(Cond2)))))
6324 Opc = Instruction::And;
6325 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)),
6326 m_OneUse(m_Value(Cond2)))))
6327 Opc = Instruction::Or;
6331 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) ||
6332 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) )
6335 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump());
6338 auto *InsertBefore = std::next(Function::iterator(BB))
6339 .getNodePtrUnchecked();
6340 auto TmpBB = BasicBlock::Create(BB.getContext(),
6341 BB.getName() + ".cond.split",
6342 BB.getParent(), InsertBefore);
6344 // Update original basic block by using the first condition directly by the
6345 // branch instruction and removing the no longer needed and/or instruction.
6346 Br1->setCondition(Cond1);
6347 LogicOp->eraseFromParent();
6349 // Depending on the conditon we have to either replace the true or the false
6350 // successor of the original branch instruction.
6351 if (Opc == Instruction::And)
6352 Br1->setSuccessor(0, TmpBB);
6354 Br1->setSuccessor(1, TmpBB);
6356 // Fill in the new basic block.
6357 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB);
6358 if (auto *I = dyn_cast<Instruction>(Cond2)) {
6359 I->removeFromParent();
6360 I->insertBefore(Br2);
6363 // Update PHI nodes in both successors. The original BB needs to be
6364 // replaced in one succesor's PHI nodes, because the branch comes now from
6365 // the newly generated BB (NewBB). In the other successor we need to add one
6366 // incoming edge to the PHI nodes, because both branch instructions target
6367 // now the same successor. Depending on the original branch condition
6368 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that
6369 // we perfrom the correct update for the PHI nodes.
6370 // This doesn't change the successor order of the just created branch
6371 // instruction (or any other instruction).
6372 if (Opc == Instruction::Or)
6373 std::swap(TBB, FBB);
6375 // Replace the old BB with the new BB.
6376 for (auto &I : *TBB) {
6377 PHINode *PN = dyn_cast<PHINode>(&I);
6381 while ((i = PN->getBasicBlockIndex(&BB)) >= 0)
6382 PN->setIncomingBlock(i, TmpBB);
6385 // Add another incoming edge form the new BB.
6386 for (auto &I : *FBB) {
6387 PHINode *PN = dyn_cast<PHINode>(&I);
6390 auto *Val = PN->getIncomingValueForBlock(&BB);
6391 PN->addIncoming(Val, TmpBB);
6394 // Update the branch weights (from SelectionDAGBuilder::
6395 // FindMergedConditions).
6396 if (Opc == Instruction::Or) {
6397 // Codegen X | Y as:
6406 // We have flexibility in setting Prob for BB1 and Prob for NewBB.
6407 // The requirement is that
6408 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB)
6409 // = TrueProb for orignal BB.
6410 // Assuming the orignal weights are A and B, one choice is to set BB1's
6411 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice
6413 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB.
6414 // Another choice is to assume TrueProb for BB1 equals to TrueProb for
6415 // TmpBB, but the math is more complicated.
6416 uint64_t TrueWeight, FalseWeight;
6417 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6418 uint64_t NewTrueWeight = TrueWeight;
6419 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight;
6420 scaleWeights(NewTrueWeight, NewFalseWeight);
6421 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6422 .createBranchWeights(TrueWeight, FalseWeight));
6424 NewTrueWeight = TrueWeight;
6425 NewFalseWeight = 2 * FalseWeight;
6426 scaleWeights(NewTrueWeight, NewFalseWeight);
6427 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6428 .createBranchWeights(TrueWeight, FalseWeight));
6431 // Codegen X & Y as:
6439 // This requires creation of TmpBB after CurBB.
6441 // We have flexibility in setting Prob for BB1 and Prob for TmpBB.
6442 // The requirement is that
6443 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB)
6444 // = FalseProb for orignal BB.
6445 // Assuming the orignal weights are A and B, one choice is to set BB1's
6446 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice
6448 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB.
6449 uint64_t TrueWeight, FalseWeight;
6450 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) {
6451 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight;
6452 uint64_t NewFalseWeight = FalseWeight;
6453 scaleWeights(NewTrueWeight, NewFalseWeight);
6454 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext())
6455 .createBranchWeights(TrueWeight, FalseWeight));
6457 NewTrueWeight = 2 * TrueWeight;
6458 NewFalseWeight = FalseWeight;
6459 scaleWeights(NewTrueWeight, NewFalseWeight);
6460 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext())
6461 .createBranchWeights(TrueWeight, FalseWeight));
6465 // Note: No point in getting fancy here, since the DT info is never
6466 // available to CodeGenPrepare.
6471 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();
6477 void CodeGenPrepare::stripInvariantGroupMetadata(Instruction &I) {
6478 if (auto *InvariantMD = I.getMetadata(LLVMContext::MD_invariant_group))
6479 I.dropUnknownNonDebugMetadata(InvariantMD->getMetadataID());